Artificial intelligence-based molecular processing method and apparatus, electronic equipment, and computer program

Abstract

The present application provides an artificial intelligence-based molecular processing method and its device, electronic device, computer-readable storage medium and computer program product, the method including: obtaining a three-dimensional structure of a target molecule; calling a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule to obtain an energy error of the target molecule, where the neural network model is trained by fitting the energy error of the sample molecule, and the energy error of the sample molecule refers to the difference between the calculation results obtained when calculating the energy of the sample molecule according to two kinds of energy calculation mechanisms respectively; performing a first energy calculation processing on the three-dimensional structure of the target molecule to obtain a first energy of the target molecule; and performing an error correction processing on the first energy of the target molecule based on the energy error of the target molecule to obtain a second energy of the target molecule.

Description

[Technical Field]

[0001] (Cross-reference to related applications) This application is filed based on the Chinese patent application filed with the China National Patent Office on 16 August 2022, application number 202210980553.2, claiming priority from the Chinese patent application, and all contents of the Chinese patent application are incorporated into this application by reference.

[0002] This application relates to artificial intelligence technology, and more particularly to molecular processing methods and apparatus based on artificial intelligence, electronic devices, computer-readable storage media, and computer program products. [Background technology]

[0003] Artificial intelligence (AI) is a comprehensive field of computer science that studies the design principles and implementation methods of various intelligent machines, enabling them to perceive, reason, and make decisions. AI technology is a broad discipline encompassing various fields such as natural language processing and machine learning / deep learning. With technological advancements, AI technology is being applied to more fields and is playing an increasingly important role.

[0004] In the research and development processes for new drugs and materials, as shown in Figure 1, after target identification and validation are completed, candidate drug compounds need to be screened. The screening process typically requires calculating molecular energy in scenarios such as molecular property calculations. The calculation results are useful for drug researchers to analyze molecular properties, the ability of molecules to bind to protein pockets, and so on. Related technologies usually use computational chemistry to obtain molecular energy, but these methods cannot account for both the accuracy and speed of molecular energy acquisition. [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] Embodiments of the present application provide an artificial intelligence-based molecular processing method, an apparatus therefor, an electronic device, a computer-readable storage medium, and a computer program product that can simultaneously improve the accuracy and speed of molecular energy calculation.

Means for Solving the Problems

[0006] The technical solution of the embodiments of the present application is realized as follows.

[0007] Embodiments of the present application provide an artificial intelligence-based molecular processing method executed by an electronic device, the method comprising: obtaining a three-dimensional structure of a target molecule; calling a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule to obtain an energy error of the target molecule, wherein the neural network model is trained by fitting energy errors of sample molecules, and the energy error of the sample molecule refers to the difference between calculation results obtained when calculating the energy of the sample molecule according to two types of energy calculation mechanisms respectively, the two types of energy calculation mechanisms include a first energy calculation process and a second energy calculation process, the accuracy of the first energy calculation process is lower than that of the second energy calculation process, and the speed of the first energy calculation process is faster than that of the second energy calculation process; performing the first energy calculation process on the three-dimensional structure of the target molecule to obtain a first energy of the target molecule; and performing error correction processing on the first energy of the target molecule based on the energy error of the target molecule to obtain a second energy of the target molecule. The embodiments of the present application provide an artificial intelligence-based molecular processing apparatus, the apparatus comprising:

[0008] an acquisition module configured to acquire a three-dimensional structure of a target molecule; ​A neural network module configured to call a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule and obtain the energy error of the target molecule, where the neural network model is trained by fitting the energy errors of sample molecules, and the energy errors of the sample molecules refer to the differences in the calculation results obtained when calculating the energies of the sample molecules according to two types of energy calculation mechanisms respectively. The two types of energy calculation mechanisms include a first energy calculation process and a second energy calculation process. The accuracy of the first energy calculation process is lower than that of the second energy calculation process, and the speed of the first energy calculation process is faster than that of the second energy calculation process. A neural network module, A calculation module configured to perform the first energy calculation process on the three-dimensional structure of the target molecule and obtain the first energy of the target molecule, A correction module configured to perform error correction processing on the first energy of the target molecule based on the energy error of the target molecule and obtain the second energy of the target molecule.

[0009] Embodiments of the present application A memory storing computer-executable instructions, A processor that executes the computer-executable instructions stored in the memory to implement a molecule processing method based on artificial intelligence according to embodiments of the present application. An electronic device comprising the processor and the memory is provided.

[0010] Embodiments of the present application provide a computer-readable storage medium storing computer-executable instructions for causing a processor to implement a molecule processing method based on artificial intelligence according to embodiments of the present application when the instructions are executed by the processor.

[0011] The embodiment of the present application provides a computer program product that, when executed by a processor, includes a computer program or computer executable instructions for enabling the processor to implement the artificial intelligence-based molecular processing method according to the embodiment of the present application. [Effects of the Invention]

[0012] The embodiments of this application have the following beneficial effects.

[0013] The first energy of the target molecule is obtained by performing a first energy calculation on the three-dimensional structure of the target molecule. Since the first energy calculation is faster than the second energy calculation, the energy calculation speed is increased, and an energy error prediction process is performed on the three-dimensional structure of the target molecule using a neural network model to obtain the energy error of the target molecule. An error correction process is then performed on the first energy calculated based on the energy error to obtain the second energy of the target molecule. Since the energy error can represent the difference in calculation results between the highly accurate second energy calculation and the less accurate first energy calculation, the accuracy of the second energy can be improved by correcting the first energy calculated using the energy error predicted by deep learning. [Brief explanation of the drawing]

[0014] [Figure 1] This is an illustrative flowchart of pharmaceutical research and development according to the embodiment of the present invention. [Figure 2] This is a schematic diagram of the architecture of an artificial intelligence-based molecular processing system according to an embodiment of the present invention. [Figure 3] This is an illustrative structural diagram of an electronic device according to an embodiment of the present invention. [Figure 4A] This is an illustrative flowchart of a molecular processing method based on artificial intelligence according to the embodiments of the present invention. [Figure 4B] This is an illustrative flowchart of a molecular processing method based on artificial intelligence according to the embodiments of the present invention. [Figure 4C]This is an illustrative flowchart of a molecular processing method based on artificial intelligence according to the embodiments of the present invention. [Figure 5] This is an illustrative structural diagram of a deep quantum chemical model according to an embodiment of the present invention. [Figure 6] This is a comparative diagram of the effects of the first dataset according to the embodiment of the present application. [Figure 7] This is a comparative diagram of the effects of the second dataset according to the embodiment of the present application. [Modes for carrying out the invention]

[0015] To further clarify the purpose, technical solutions and advantages of this application, the present application will be described in more detail below with reference to the drawings, and the embodiments described therein should not be understood as limitations to this application, and all other embodiments that can be obtained by those skilled in the art without creative effort are included within the scope of protection of the present invention.

[0016] In the following description, the term “several embodiments” refers to a subset of all possible embodiments, and as can be understood, “several embodiments” may be the same subset or different subsets of all possible embodiments, and these can be combined with one another without conflict.

[0017] The terms “first / second / third” as used in the following description are merely for distinguishing similar subjects and do not represent a specific order of subjects. Understandably, “first / second / third” may, in some cases, be interchangeable with a specific order or sequence, thereby allowing the embodiments of the present application described herein to be carried out in an order other than that shown or described.

[0018] Unless otherwise specified, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art. The terms used herein are for illustrative purposes only and are not intended to limit the application.

[0019] Before describing the embodiments of this application in further detail, the nouns and terms referred to in the embodiments of this application are explained below, with the following interpretations.

[0020] 1) Computational chemistry is a branch of theoretical chemistry that primarily aims to calculate molecular properties, such as total energy, dipole moment, quadrupole moment, vibrational frequency, and reactivity, using effective mathematical approximations and computer programs, and is used to interpret certain specific chemical problems.

[0021] 2) Molecular energy includes the kinetic energy and potential energy of molecules. Molecular kinetic energy refers to the energy due to the motion of a molecule, while molecular potential energy refers to the potential energy that a molecule possesses, which is generated by molecular forces and determined by its relative position.

[0022] 3) Regarding semi-empirical quantum mechanical methods, while theoretically quantum mechanical methods offer high accuracy, their computational cost is extremely high when dealing with biomolecular systems, making them unsuitable for practical applications. Semi-empirical quantum mechanical methods, which incorporate empirical parameters, are approximate methods with a trade-off relationship between time and accuracy. They can be used for scoring and estimating ligand-protein affinity and have significant importance in computer-aided drug design.

[0023] 4) Delta learning refers to a learning system that can continuously learn new knowledge from new samples while retaining most of the previously learned knowledge. Delta learning is very similar to the learning mode of humans.

[0024] 5) The first-principles method (AB Initio Method) refers to a quantum chemical calculation method that directly solves the Schrödinger equation based on the fundamental principles of quantum mechanics. The characteristics of the first-principles method are the absence of empirical parameters and the fact that the system is not overly simplified. Calculations are performed using essentially the same method for various different chemical systems.

[0025] 6) Density Functional Theory is a method for studying the electronic structure of many-body electron systems. Density Functional Theory has a wide range of applications in physics and chemistry, and is particularly used to study the properties of molecules and aggregated states. It is one of the most widely used methods in the fields of aggregated state physics, computational materials science, and computational chemistry.

[0026] Related technologies for calculating molecular energy include two types: computational chemistry-based methods and deep learning-based methods. Computational chemistry methods include quantum mechanics methods and molecular mechanics (MM) methods.

[0027] Quantum mechanics methods include the first-principles method (AB Initio Method) and semi-empirical quantum mechanical methods. The first-principles method is based on the first principles of quantum chemistry and solves the Schrödinger equation using exact approximations. The first-principles method includes methods based on wave functions and methods based on density functionals. A representative method based on wave functions is the method based on the Hartree-Fock equation, and a representative method based on density functionals is the method based on density functional theory.

[0028] In molecular mechanics, the influence of various forms of interaction forces on the potential energy of molecules is typically described using a molecular force field. This is a simplified model of molecular structure that does not involve calculating electron interactions.

[0029] In the deep learning method, molecular energy is primarily predicted using a deep learning model. A small dataset is used as the training sample set, and the deep learning model is trained based on this set of training samples. As a result, the deep learning model learns the features of atomic structures and can predict molecular energy based on the features obtained through extraction.

[0030] The first-principles method of quantum mechanics offers high computational accuracy but requires a large amount of computation time, making large-scale calculations difficult. The semi-empirical quantum mechanics method, on the other hand, uses semi-empirical parameters instead of molecular integrals, which speeds up calculations but sacrifices computational accuracy.

[0031] Methods based on molecular mechanics force fields can achieve faster computation speeds, but they are also the least accurate methods.

[0032] Deep learning methods do not fully utilize atomic features, support a limited number of atomic types, have a poor ability to capture long-range interactions between molecules, and do not explore predictions for the energy of different molecular conformations. Simple deep learning methods need to improve their accuracy in molecular energy prediction.

[0033] To solve the above problems, the embodiments of the present application provide an artificial intelligence-based molecular processing method and apparatus, electronic device, computer-readable storage medium, and computer program product that can simultaneously improve the accuracy and speed of molecular energy calculations.

[0034] The molecular processing method based on artificial intelligence according to the embodiment of the present invention may be implemented independently by a terminal / server or collaboratively by a terminal and a server. For example, the terminal may independently perform the molecular processing method based on artificial intelligence described later, or the terminal may send an energy evaluation request for a target molecule to the server, and the server may execute the molecular processing method based on artificial intelligence based on the received energy evaluation request for the target molecule to obtain the three-dimensional structure of the target molecule, call a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule to obtain the energy error of the target molecule, perform a first energy calculation process on the three-dimensional structure of the target molecule to obtain the first energy of the target molecule, and perform error correction processing on the first energy of the target molecule based on the energy error of the target molecule to obtain the second energy of the target molecule. As a result, the researcher can perform subsequent analytical research based on the second energy of the target molecule, for example, to determine the binding ability of the target molecule to protein pockets based on the second energy of the target molecule, and to screen candidate drug compounds through the binding ability of the target molecule to protein pockets.

[0035] The electronic equipment for molecular processing according to the embodiments of this application may be various types of terminal devices or servers, where a server may be an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services, and a terminal may be, but is not limited to, a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc. The terminal and server may be connected directly or indirectly by wired or wireless communication methods, and this is not limited in this application.

[0036] Taking servers as an example, this could be a server cluster located in the cloud, which opens up artificial intelligence services (AIaaS: AI as a Service) to users. The AIaaS platform divides several common AI services and provides them as independent or packaged services on the cloud. This service mode is similar to an AI thememol, where all users can access and use one or more artificial intelligence services provided by the AIaaS platform through an application programming interface.

[0037] For example, the artificial intelligence service here may be a molecular processing service, that is, a molecular processing program according to the embodiment of this application is encapsulated on a cloud server. The user invokes the molecular processing service within the cloud service via a terminal (where a client such as a compound screening client is running), causing the server located on the cloud to invoke the encapsulated molecular processing program to obtain the three-dimensional structure of the target molecule, invoke a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule to obtain the energy error of the target molecule, perform first energy calculation processing on the three-dimensional structure of the target molecule to obtain the first energy of the target molecule, and perform error correction processing on the first energy of the target molecule based on the energy error of the target molecule to obtain the second energy of the target molecule. As a result, the researcher can perform subsequent analytical research based on the second energy of the target molecule, for example, to determine the binding ability of the target molecule to protein pockets based on the second energy of the target molecule, and to screen candidate drug compounds through the binding ability of the target molecule to protein pockets.

[0038] Referring to Figure 2, which is a schematic diagram of the architecture of an artificial intelligence-based molecular processing system according to an embodiment of the present invention, the terminal 400 is connected to the server 200 via the network 300, and the network 300 may be a wide area network, a local area network, or a combination of both.

[0039] Terminal 400 (where a client such as a compound screening client is running) can be used to obtain an energy evaluation request for a target molecule. For example, when a researcher inputs a target molecule through the input interface of terminal 400, an energy evaluation request for the target molecule is automatically generated. Terminal 400 sends the energy prediction request for the target molecule to server 200, which obtains the three-dimensional structure of the target molecule, invokes a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule to obtain the energy error of the target molecule, performs a first energy calculation processing on the three-dimensional structure of the target molecule to obtain the first energy of the target molecule, performs error correction processing on the first energy of the target molecule based on the energy error of the target molecule to obtain the second energy of the target molecule, and returns the second energy of the target molecule to terminal 400. As a result, the researcher can perform subsequent analytical studies based on the second energy of the target molecule, for example, to determine the binding ability of the target molecule to protein pockets based on the second energy of the target molecule, and to screen candidate drug compounds through the binding ability of the target molecule to protein pockets.

[0040] In some embodiments, a molecular processing plugin can be embedded in a client running on a terminal, allowing the client to implement an artificial intelligence-based molecular processing method locally. For example, terminal 400, after obtaining an energy evaluation request for a target molecule, invokes the molecular processing plugin to implement an artificial intelligence-based molecular processing method, obtains the three-dimensional structure of the target molecule, invokes a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule, obtains the energy error of the target molecule, performs a first energy calculation process on the three-dimensional structure of the target molecule to obtain the first energy of the target molecule, and performs error correction processing on the first energy of the target molecule based on the energy error of the target molecule to obtain the second energy of the target molecule. As a result, researchers can perform subsequent analytical studies based on the second energy of the target molecule, for example, by determining the binding ability of the target molecule to protein pockets based on the second energy of the target molecule, and screening of candidate drug compounds can be performed through the binding ability of the target molecule to protein pockets.

[0041] In some embodiments, after the terminal 400 obtains an energy evaluation request for a target molecule, it invokes the molecular processing interface of the server 200 (which can be provided as a cloud service, i.e., as a molecular processing service). The server 200 obtains the three-dimensional structure of the target molecule, invokes a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule, obtains the energy error of the target molecule, performs first energy calculation processing on the three-dimensional structure of the target molecule to obtain the first energy of the target molecule, and performs error correction processing on the first energy of the target molecule based on the energy error of the target molecule to obtain the second energy of the target molecule. As a result, the researcher can perform subsequent analytical studies based on the second energy of the target molecule. For example, based on the second energy of the target molecule, the ability of the target molecule to bind to a protein pocket can be determined, and candidate drug compounds can be screened based on the ability of the target molecule to bind to a protein pocket.

[0042] Referring to Figure 3, which is an exemplary structural diagram of an electronic device according to an embodiment of the present invention, the terminal 400 shown in Figure 3 comprises at least one processor 410, memory 450, at least one network interface 420, and a user interface 430. Each component within the terminal 400 is coupled by a bus system 440. Understandably, the bus system 440 is used to enable connection and communication between these components. In addition to the data bus, the bus system 440 further includes a power bus, a control bus, and a status signal bus. However, for clarity in the explanation, in Figure 3, the various buses are labeled as the bus system 440.

[0043] The processor 410 may be an integrated circuit chip having signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, where the general-purpose processor may be a microprocessor or any conventional processor.

[0044] The user interface 430 includes one or more output devices 431 that enable the display of media content, and each output device 431 includes one or more speakers and / or one or more visual display screens. The user interface 430 further includes one or more input devices 432, each input device 432 includes user interface components that assist user input, such as a keyboard, mouse, microphone, touchscreen display, camera, and other input buttons and controls.

[0045] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid memory, hard disk drivers, optical disk drivers, and the like. The memory 450 optionally includes one or more storage devices located physically away from the processor 410.

[0046] The memory 450 may include volatile memory or non-volatile memory, or it may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random-access memory (RAM). The memory 450 described in this embodiment is intended to include any suitable type of memory.

[0047] In some embodiments, the memory 450 can store data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0048] Operating System 451 includes system programs for handling various basic system services and performing hardware-related tasks, such as a framework layer, core library layer, and driver layer, which are used to implement various basic services and handle hardware-based tasks.

[0049] The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420, and exemplary network interfaces 420 include Bluetooth® technology, Wireless Fidelity (WiFi), and Universal Serial Bus (USB).

[0050] The display module 453 is used to enable information to be displayed via one or more output devices 431 (e.g., a display screen, a speaker, etc.) associated with the user interface 430 (e.g., a user interface for operating peripheral devices and displaying content and information).

[0051] The input processing module 454 is used to detect one or more user inputs or interactions from one or more input devices 432 and to convert the detected inputs or interactions.

[0052] In some embodiments, the molecular processing apparatus based on artificial intelligence according to the embodiments of the present invention can be implemented in software form, and Figure 3 shows the molecular processing apparatus based on artificial intelligence 455 stored in memory 450. The molecular processing apparatus based on artificial intelligence 455 may be software in the form of a program or plug-in, and includes software modules such as an acquisition module 4551, a neural network module 4552, a computation module 4553, a correction module 4554, and a training module 4555. Since these modules are logical, they can be arbitrarily combined or further divided based on the functions to be implemented. The functions of each module will be described later.

[0053] As described above, the molecular processing method based on artificial intelligence according to the embodiments of this application can be carried out by various types of electronic devices.

[0054] The following describes a molecular processing method based on artificial intelligence according to an embodiment of the present application. Similarly, the electronic device that implements the molecular processing method based on artificial intelligence according to an embodiment of the present application may be a terminal. Therefore, the entity that performs each step will not be described repeatedly below.

[0055] Referring to Figure 4A, Figure 4A is an exemplary flowchart of an artificial intelligence-based molecular processing method according to an embodiment of the present invention, and will be explained with reference to steps 101 to 104 shown in Figure 4A.

[0056] In step 101, the three-dimensional structure of the target molecule is obtained.

[0057] For example, the target molecule has multiple molecular conformations, each of which has a corresponding three-dimensional structure, and the target molecule contains at least one atom. In the case of a target molecule with multiple atoms, the three-dimensional structure is a conformation consisting of multiple atoms and at least one chemical bond.

[0058] In step 102, a neural network model is invoked to perform energy error prediction processing on the three-dimensional structure of the target molecule, and the energy error of the target molecule is obtained.

[0059] As an example, a neural network model is used to extract features from the three-dimensional structure of a target molecule, obtain energy error features, and map these energy error features to the difference in calculation results obtained when calculating the energy of the target molecule using two given energy calculation mechanisms. The neural network model performs fitting based on feature mapping to obtain the energy error, and in practice, it is not necessary to perform energy calculations on the target molecule using the two given energy calculation mechanisms. This is because the functionality that the neural network model can achieve depends on the training method. The neural network model is trained by fitting the energy error of a sample molecule, and the energy error of a sample molecule refers to the difference in calculation results obtained when calculating the energy of the sample molecule according to the two energy calculation mechanisms, respectively. The energy calculation mechanism represents a method for performing energy calculations on a given molecule, and the energy calculation method may be an energy calculation process based on computational chemistry, which includes quantum mechanical methods and molecular mechanical methods. The quantum mechanical methods include first-principles methods (e.g., calculations using the Psi4 tool) and semi-empirical quantum mechanical methods (e.g., calculations using the GFN2-xTB program and the GFN-xTB program). The two types of energy calculation mechanisms correspond to a first energy calculation process and a second energy calculation process, respectively. The accuracy of the first energy calculation process is less than that of the second energy calculation process, and the speed of the first energy calculation process is faster than that of the second energy calculation process.

[0060] In step 103, the first energy calculation process is performed on the three-dimensional structure of the target molecule to obtain the first energy of the target molecule.

[0061] For example, the first and second energy calculation processes may be energy calculation processes in computational chemistry, and the energy calculation processes in computational chemistry include quantum mechanical methods and molecular mechanical methods. The quantum mechanical methods include first-principles methods (e.g., calculations using the Psi4 tool) and semi-empirical quantum mechanical methods (e.g., using the GFN2-xTB program and the GFN-xTB program), the first energy calculation process may be a semi-empirical quantum mechanical method, and the second energy calculation process may be a first-principles method, the first-principles method being a method that solves the Schrödinger equation using exact approximations based on the first principles of quantum chemistry. The first-principles methods include wave function-based methods and density functional-based methods, a typical wave function-based method being the method based on the Hartree-Fock equation, and a typical density functional-based method being the method based on density functional theory.

[0062] In step 104, an error correction process is performed on the first energy of the target molecule based on the energy error of the target molecule to obtain the second energy of the target molecule.

[0063] For example, error correction processing refers to adding or subtracting the first energy and the energy error. If the energy error represents the difference in calculation results between the first energy calculation and the second energy calculation, the error correction processing involves subtracting the first energy and the energy error to obtain the second energy of the target nutrient. If the energy error represents the difference in calculation results between the second energy calculation and the first energy calculation, the error correction processing involves adding the first energy and the energy error to obtain the second energy of the target nutrient.

[0064] Referring to Figure 4B in some embodiments, Figure 4B is an exemplary flowchart of an artificial intelligence-based molecular processing method according to an embodiment of the present invention. In step 102, the step of calling a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule and obtaining the energy error of the target molecule can be achieved through steps 1021 to 1022 shown in Figure 4B.

[0065] In step 1021, a neural network model is used to perform feature extraction on the three-dimensional structure and obtain the energy error features of the target molecule.

[0066] In step 1022, a fully connected neural network model is used to perform a full-connection process on the energy error features to obtain the energy error of the target molecule.

[0067] The process of extracting energy error features of a three-dimensional structure using artificial intelligence methods and performing the feature extraction process in step 1021 will be explained in detail in the following sections. Based on the energy error features, the energy error of the target molecule can be predicted, and the difference between the calculated molecular energy results of the target molecule, calculated using two types of energy calculation processes, can be intelligently determined.

[0068] In some embodiments, the neural network model includes N cascaded feature networks, and the step of extracting features from a three-dimensional structure using the neural network model of step 1021 to obtain the energy error features of the target molecule can be achieved through the following steps A to C.

[0069] In step A, an initial feature extraction process is performed on each atom in the three-dimensional structure to obtain the initial characteristics of each atom.

[0070] For example, if the target molecule E contains three atoms (atom A, atom B, and atom C) and one chemical bond (atom A and atom B are connected by a chemical bond), then an initial feature extraction process is performed on each atom to obtain the initial attribute features of each atom in the three-dimensional structure. These initial attribute features represent the attribute information of the atom (at least one of the following: atom type, atom properties, etc.). Then, the initial coordinate features of each atom in the three-dimensional structure are obtained. These initial coordinate features represent the position information of the atom (the position information refers to the three-dimensional coordinate of each atom, and by establishing a three-dimensional coordinate system with any one of the three atoms as the origin, the three-dimensional coordinates of the other two atoms can be determined). The following process is then performed on each atom in the three-dimensional structure, for example, the following process is performed on atom A. At least one other atom (i.e., atom B and atom C) excluding atom A in the three-dimensional structure is obtained, and initial relationship features are obtained between atom A and each of the other atoms. These initial relationship features represent the bonding relationships between atom A and the other atoms. The initial relational features represent the bonding relationship between atom A and atom B, and between atom A and atom C. There are two types of bonding relationships: bonded relationships and unbonded relationships. If two atoms are connected by a chemical bond, they have a bonded relationship. If two atoms are not connected by a chemical bond, they have an unbonded relationship. The initial attribute features, initial coordinate features, and initial relational features of each atom are defined as the initial features of each atom.

[0071] By using initial attribute features, initial relational features, and initial coordinate features as initial features, it is possible to represent not only the type attributes of atoms but also their relative positions and the connection status due to chemical bonds, thereby effectively improving the expressive power of the initial features.

[0072] In step B, if the value of n is 1 ≤ n ≤ N-1, the nth feature network among the N cascaded feature networks is used to perform an nth feature extraction process on the input of the nth feature network to obtain the nth feature of each atom, and the nth feature is transmitted to the (n+1)th feature network.

[0073] For example, if the range of N satisfies 2 ≤ N, the value of n is an integer increasing from 1, and the range of n satisfies 1 ≤ n ≤ N, and the value of n is 1, then the input to the nth feature network is the initial features of each atom. If the value of n is 2 ≤ n ≤ N, then the input to the nth feature network is the (n-1)th feature of each atom output by the (n-1)th feature network.

[0074] As an example, assuming that the value of N is 3, we still perform the following process using atom A as an example. Using the first feature network, we perform a first feature extraction process on the initial features of atom A to obtain the first feature of atom A, transmit the first feature to the second feature network and continue the second feature extraction process. Using the second feature network, we perform a second feature extraction process on the first feature of atom A to obtain the second feature of atom A.

[0075] In step C, if the value of n is N, attribute feature extraction is performed on the (n-th)th feature of each atom using the n-th feature network to obtain the n-th attribute feature of each atom, coordinate feature extraction is performed on the (n-th)th feature of each atom using the n-th feature network to obtain the n-th coordinate feature of each atom, and an energy error feature is formed using the n-th attribute feature and n-th coordinate feature of each atom.

[0076] As an example, assuming that the value of N is 3, the following process is still performed using atom A as an example. A third-feature network is used to extract third-attribute features from the second feature of atom A to obtain the third-attribute features of atom A. A third-feature network is used to extract third-coordinate features from the second feature of atom A to obtain the third-coordinate features of atom A. The third-attribute features and the third-coordinate features of atom A are then used as energy error features.

[0077] According to the embodiment of the present invention, by obtaining the coordinate features and attribute features of each atom using a sequential substitution method, and forming an energy error feature using the coordinate features and attribute features of multiple atoms, the feature representation capability of the energy error feature can be effectively improved, thereby improving the accuracy of subsequent predicted energy errors.

[0078] In some embodiments, the step of using the nth feature network among the N cascaded feature networks of step B to perform an nth feature extraction process on the input of the nth feature network to obtain the nth feature of each atom can be achieved by performing the following steps B1 to B3 for each atom (let's take atom A in target molecule E as an example) using the nth feature network.

[0079] In step B1, we obtain the other atoms excluding the atoms in the three-dimensional structure, and still using the target molecule E as an example, the other atoms are atom B and atom C, excluding atom A.

[0080] In step B2, a first mapping process is performed on the (n-1)th feature of an atom and the (n-1)th feature of each other atom to obtain the nth association feature of the atom corresponding to each other atom.

[0081] The following process is performed for each other atom (atom B will be used as an example later): The (n-th)th coordinate feature of an atom is extracted from the (n-th)th feature of an atom, the (n-th)th coordinate feature of each other atom is extracted from the (n-th)th feature of each other atom, the (n-th)th attribute feature of an atom is extracted from the (n-th)th feature of an atom, the (n-th)th attribute feature of each other atom is extracted from the (n-th)th feature of each other atom, the (n-th)th relation feature of an atom is extracted from the (n-th)th feature of an atom, and the following process is performed for each other atom (atom B will be used as an example later): The (n-th)th relation feature of an atom with respect to other atoms is extracted from the (n-th)th relation feature of an atom, the first feature distance between the (n-th)th coordinate feature of an atom and the (n-th)th coordinate feature of other atoms is obtained, and the first fusion process is performed on the square of the first feature distance, the (n-th)th attribute feature of an atom, the (n-th)th attribute feature of other atoms, and the (n-th)th relation feature of an atom with respect to other atoms to obtain the nth association feature of an atom corresponding to other atoms.

[0082] According to the embodiments of the present application, since the characteristic distance between any two atoms and the bonding relationship between any two atoms are considered as associated features, the associated features can learn the global information of the three-dimensional structure and improve the global information learning ability of the neural network model.

[0083] As an example, assuming that the value of N is 3, still taking atom A and another atom B as examples, the second coordinate feature of atom A is extracted from the second feature of atom A, the second coordinate feature of atom B is extracted from the second feature of atom B, the second attribute feature of the atom is extracted from the second feature of atom A, and the second attribute feature of atom B is extracted from the second feature of atom B. The (n - 1)-th relationship feature is the initial relationship feature, that is, the relationship feature does not change by successive substitution of features, and the second relationship feature (the initial relationship feature of atom A) of atom A is extracted from the second feature of atom A, and the second relationship feature (the initial relationship feature of atom B) of the atom is extracted from the second feature of atom B.

[0084] As an example, referring to formula (1), subsequent processing is performed on another atom B.

[0085]

Number

Number

Number

[0086] In step B3, a second mapping process is performed on the (n-1)th feature of an atom and the nth association feature of each other atom to obtain the nth feature of an atom.

[0087] Addition is performed on the nth association features of the atom corresponding to multiple other atoms to obtain the nth association feature of the atom. A second fusion process is performed on the (n-th)th attribute feature of the atom and the nth association feature of the atom to obtain the nth attribute feature of the atom. The first feature difference between the (n-th)th coordinate feature of the atom and the (n-th)th coordinate feature of each other atom is obtained. Linear mapping is performed on the nth association features of the atom corresponding to each other atom to obtain the weight of each other atom. Based on the weight of each other atom, a weighted average is performed on the first feature difference of multiple other atoms to obtain the weighted average result corresponding to the atom. Addition is performed on the weighted average result of the atom and the (n-th)th coordinate feature of the atom to obtain the nth coordinate feature of the atom. The initial relationship feature of the atom is used as the nth relationship feature of the atom. The nth feature of the atom is formed using the nth relationship feature of the atom, the nth attribute feature of the atom, and the nth coordinate feature of the atom.

[0088] In the embodiment of the present invention, the feature distance between any two atoms and the bonding relationship between any two atoms are considered as the nth feature. Therefore, the nth feature can learn global information of the three-dimensional structure, thereby improving the global information learning ability of the neural network model.

[0089] As an example, by referring to equation (2), an addition operation is performed on the nth association feature of atom A corresponding to multiple other atoms (other atoms B and other atoms C) to obtain the nth association feature of atom A.

[0090]

number

[0091] As an example, referring to equation (3), a second fusion process is performed on the (n-1)th attribute feature and the nth association feature of atom A to obtain the nth attribute feature of atom A.

[0092]

number

number

[0093] As an example, the first feature difference is obtained between the (n-1)th coordinate feature of atom A and the (n-1)th coordinate feature of other atoms (other atoms B and other atoms C). A weighted average is performed on the first feature differences of multiple other atoms (other atoms B and other atoms C), using the linear mapping result of the nth association feature of atom A corresponding to each other atom (other atoms B and other atoms C) as a weight to obtain the weighted average result for atom A. This weighted average result for atom A is then added to the (n-1)th coordinate feature of atom A to obtain the nth coordinate feature of atom A. The above processing for atom A and the other atoms can be referenced from equation (4).

[0094]

number

number

[0095] As an example, let the initial relational feature of atom A be the nth relational feature of atom A, and let the nth relational feature a of atom A correspond to each other atom. ij , the nth attribute characteristic h of atom A i n The nth coordinate feature of atom A is used to form the nth feature of the atom.

[0096] In some embodiments, the step of obtaining the nth attribute feature of each atom by performing attribute feature extraction on the (n-th)th feature of each atom using the nth feature network of step C is realized through the following technical solution: Extract the (n-th)th coordinate feature of the atom, the (n-th)th attribute feature of the atom, and the (n-th)th relation feature of the atom from the (n-th)th feature of each atom, and perform the following process for each atom. Obtain the other atoms excluding the atoms in the three-dimensional structure, and perform the following process for each of the other atoms. The (n-th)th relational features of an atom are extracted from the (n-th)th relational features of an atom to other atoms, the second feature distance is obtained between the (n-th)th coordinate features of an atom and the (n-th)th coordinate features of other atoms, the square of the second feature distance, the (n-th)th attribute features of an atom, the (n-th)th attribute features of other atoms, and the (n-th)th relational features of an atom to other atoms are subjected to a first fusion process to obtain the nth relational feature of an atom corresponding to other atoms, an addition process is performed on the nth relational features of an atom corresponding to multiple other atoms to obtain the nth relational feature of an atom, and the second fusion process is performed on the (n-th)th attribute features of an atom and the nth relational feature of an atom to obtain the nth attribute feature of an atom.

[0097] As an example, the (n-th)-th coordinate feature, the (n-th)-th attribute feature, and the (n-th)-th relation feature of each atom are extracted from the (n-th)-th feature of each atom. Specifically, assuming that the value of N is 3, the following process is performed using atoms A, B, and C as examples: The second coordinate feature of atom A is extracted from the second feature of atom A, the second coordinate feature of atom B is extracted from the second feature of atom B, the second coordinate feature of atom C is extracted from the second feature of atom C, the second attribute feature of atom A is extracted from the second feature of atom A, the second attribute feature of atom B is extracted from the second feature of atom B, and the second attribute feature of atom C is extracted from the second feature of atom C. The (n-1)th relation feature is the initial relation feature; that is, relation features do not change with the sequential substitution of features. The second relation feature of atom A (the initial relation feature of atom A) is extracted from the second feature of atom A, the second relation feature of atom B (the initial relation feature of atom B) is extracted from the second feature of atom B, and the second relation feature of atom C (the initial relation feature of atom C) is extracted from the second feature of atom C.

[0098] As an example, referring to equation (5), for each atom (let's use atom A as an example), we obtain other atoms B and C, excluding the atoms in the three-dimensional structure, and then perform the subsequent processing on each of the other atoms (let's use other atom B as an example).

[0099]

number

number

number

[0100] As an example, referencing equation (6), an addition operation is performed on the nth association feature of atom A corresponding to multiple other atoms (other atoms B and other atoms C) to obtain the nth association feature of atom A.

[0101]

number

[0102] As an example, referring to equation (7), a second fusion process is performed on the (n-1)th attribute feature and the nth association feature of atom A to obtain the nth attribute feature of atom A.

[0103]

number

number

[0104] In step C, a specific embodiment of obtaining the nth coordinate feature of each atom by performing a coordinate feature extraction process on the (n-1)th feature of each atom using the n-feature network can be found by referring to a specific embodiment of obtaining the (n-1)th coordinate feature.

[0105] In some embodiments, referring to Figure 4C, a neural network model can be invoked to perform energy error prediction processing on the three-dimensional structure of the target molecule, and before obtaining the energy error of the target molecule, steps 105 to 109 shown in Figure 4C can be performed.

[0106] In step 105, sample molecules are obtained, and a three-dimensional structure generation process is performed on the sample molecules to obtain the three-dimensional structures of multiple sample molecules.

[0107] For example, the sample molecule can be from any molecular dataset, such as the QMugs dataset, and a structure generation process can be performed on the sample molecule to obtain multiple sample molecular structures. This structure generation process involves twisting chemical bonds, and the resulting multiple sample molecular structures include the molecular structure of the sample molecule itself and other molecular structures obtained by twisting chemical bonds. The structure generation process cannot change the types of atoms within the molecule, but it can change the coordinates and distances between atoms within the molecule.

[0108] In step 106, the label energy error for the three-dimensional structure of each sample molecule is obtained.

[0109] In some embodiments, the step of obtaining the label energy error of each sample molecular structure in step 106 can be achieved through the following technical solution. The following process is performed for each sample molecular structure: A first energy calculation process is performed on the sample molecular structure to obtain the first energy of the sample molecular structure; a second energy calculation process is performed on the sample molecular structure to obtain the second energy of the sample molecular structure; the first difference between the second energy of the sample molecule and the first energy of the sample molecule is obtained and used as the label energy error of the sample molecular structure.

[0110] As an example, the molecular energy of a molecule is calculated using a highly accurate second-energy processing method. For instance, the quantum mechanical energy (point energy) of the molecule is calculated using the DFT calculation theory level and the basis set "WB97X-D3 / def2-TZVP" and is denoted as the molecular energy. This calculated molecular energy is denoted as E_dft. Simultaneously, the molecular energy of the molecule is calculated using a high-speed first-energy processing method. For example, the semi-empirical quantum mechanical energy of the molecule is calculated using a semi-empirical quantum mechanics method and is denoted as the molecular energy. This calculated molecular energy is denoted as E_xtb, and the difference between the two, E_delta = E_dft - E_xtb, is taken as the label energy error (Label) of the molecular three-dimensional structure.

[0111] In step 107, the initialized neural network model is forward-propagated through each sample molecule's three-dimensional structure to obtain the predicted energy error for each sample molecule's three-dimensional structure.

[0112] As an example, a feature extraction process is performed on a three-dimensional structure using an initialized neural network model to obtain the energy error features of the target molecule. Then, using the initialized neural network model, a fully connected process is performed on the energy error features to obtain the energy error of the target molecule. The processing steps of the initialized neural network model can be found in the above example, but the only difference is that the parameters used in the processing steps are initialized parameters, not parameters obtained through training.

[0113] In step 108, the total loss corresponding to the neural network model is determined based on the label energy error of each sample molecular structure and the predicted energy error of each sample molecular structure.

[0114] In some embodiments, the step of determining the total loss corresponding to the neural network model based on the label energy error and the predicted energy error of each sample molecular structure in step 108 can be achieved through the following technical solutions. For each sample molecule's three-dimensional structure, a process is performed to determine the first root mean square error of the sample molecule's three-dimensional structure based on the label energy error and the predicted energy error of the sample molecule's three-dimensional structure. The other sample molecule's three-dimensional structures are obtained, and for each of these other sample molecule's three-dimensional structures, a second difference is determined between the label energy error of the sample molecule's three-dimensional structure and the label energy error of the other sample molecule's three-dimensional structure. A process is then performed to determine the third difference between the predicted energy error of the sample molecule's three-dimensional structure and the predicted energy error of the other sample molecule's three-dimensional structure. Root mean square operations are performed on the second and third differences to obtain the second root mean square error of the sample molecule's three-dimensional structure corresponding to the other sample molecule's three-dimensional structure. Addition is performed on the second root mean square errors of the sample molecule's three-dimensional structures corresponding to multiple other sample molecule's three-dimensional structures to obtain the third root mean square error of the sample molecule's three-dimensional structure. A third fusion process is performed on the first root mean square errors of multiple sample molecule's three-dimensional structures and the third root mean square errors of multiple sample molecules to obtain the total loss corresponding to the neural network model.

[0115] For example, see equation (8) for the total loss of the neural network model.

[0116]

number

number

number

number

number

number

[0117] In step 109, the neural network model is subjected to backpropagation on the total loss to obtain the parameter changes of the neural network model when the total loss converges, and the parameters of the neural network model are updated based on these parameter changes.

[0118] As an example, backpropagation can be implemented using a backpropagation algorithm, which primarily iteratively cycles through two segments (excitation propagation and weight update) and sequentially assigns values ​​until the network's response to the input reaches a predetermined target range. The learning process of the backpropagation algorithm consists of a forward propagation process and a backpropagation process. In the forward propagation process, input information is processed layer by layer from the input layer through the implicit layer and sent to the output layer. If the desired output value is not obtained at the output layer, the output and the sum of the squares of the desired error are used as the target function and passed to the backpropagation process. The partial derivatives of the target function for each neuron weight are calculated step by step, and the gradient of the target function with respect to the weight direction is constructed to provide a basis for correcting the weights. Network learning is completed in the weight correction process. When the total loss converges to the desired value, the learning of the neural network model is complete.

[0119] The following describes an exemplary application of the present embodiment in one actual application scenario.

[0120] A terminal (running a client such as a compound screening client) can be used to obtain an energy evaluation request for a target molecule. For example, when a researcher inputs a target molecule through the terminal's input interface, an energy evaluation request for the target molecule is automatically generated. The terminal sends the energy prediction request for the target molecule to the server, which obtains the three-dimensional structure of the target molecule, invokes a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule, obtains the energy error of the target molecule, performs a first energy calculation processing on the three-dimensional structure of the target molecule to obtain the first energy of the target molecule, performs error correction processing on the first energy of the target molecule based on the energy error of the target molecule to obtain the second energy of the target molecule, and returns the second energy of the target molecule to the terminal. Based on this, the researcher can perform subsequent analytical studies based on the second energy of the target molecule, for example, to determine the binding ability of the target molecule to protein pockets based on the second energy of the target molecule, and screen candidate drug compounds through the binding ability of the target molecule to protein pockets.

[0121] The molecular processing method based on artificial intelligence according to the embodiment of this application can be applied to the new drug research and development process and the materials research and development process. Taking the new drug research and development process as an example, as shown in Figure 1, in the new drug research and development process, after completing target identification and validation, it is necessary to screen candidate drug compounds. In the screening process, it is necessary to calculate the energy of molecules in scenarios such as the calculation of molecular properties and the calculation of the ability of molecules to bind to protein pockets. The calculation results help drug researchers to analyze molecular properties, the ability of molecules to bind to protein pockets, etc., and help researchers design more effective drug molecules, significantly improve the efficiency of research and development, and reduce the cost of drug research and development.

[0122] First, it is necessary to construct a training dataset for the Deep Quantum Chemistry (DeepQC) model. Based on an arbitrary dataset (for example, the QMugs dataset), an enhanced dataset is constructed. This enhanced dataset contains 660,000 molecules, and since each molecule has three molecular three-dimensional structures, there are nearly 2 million data points in total.

[0123] For each molecular structure in the enhanced dataset, the molecular energy is calculated using a highly accurate second-energy processing method. For example, the quantum mechanical energy (point energy) of the molecule is calculated using the DFT calculation theory level and the basis set "WB97X-D3 / def2-TZVP" and is denoted as the molecular energy, E_dft. Simultaneously, the molecular energy is calculated using a high-speed first-energy processing method. For example, the semi-empirical quantum mechanical energy of the molecule is calculated using a semi-empirical quantum mechanics method and is denoted as the molecular energy, E_xtb. The difference between the two, E_delta = E_dft - E_xtb, is taken as the label energy error (Label) of the molecular structure. The quantum mechanical calculation related to the second-energy processing method is implemented using the Psi4 tool, and the semi-empirical quantum mechanical calculation related to the first-energy processing method is implemented using the xTB tool. The enhanced dataset is split in an 8:1:1 ratio and used as the training, validation, and test sets for DeepQC, respectively.

[0124] The architecture of the DeepQC model is shown in Figure 5. The DeepQC model includes a neural network model related to deep learning and a first energy processing method related to computational chemistry. The neural network model may be an equivariant graph neural network. The input to the neural network model is the three-dimensional structural coordinates of the molecule, and the output to the neural network model is the predicted value of E_delta. During the training process, the total loss between the predicted value of E_delta obtained by forward propagation and the E_delta label value is calculated. Backpropagation is performed on the total loss to obtain the gradient of each network layer, and the parameters of the neural network model are updated using a self-adaptive moment estimation algorithm. For the total loss of the neural network model, please refer to equation (9).

[0125]

number

number

number

number

number

number

[0126] The training hyperparameters for the DeepQC model are shown in Table 1. [Table 1] Using a trained DeepQC model, the energy of a molecule can be predicted. The three-dimensional coordinates of the molecule are input into the DeepQC model, and the first energy E_xtb can be obtained through a first energy calculation process (e.g., processing using the GFN2-xTB program). The energy error E_delta can then be predicted using a neural network model (e.g., an equivariant graph neural network). By accumulating these two values, the final predicted energy value E_dft can be obtained.

[0127] In the research and development process for new drugs and materials, it is necessary to calculate the energy of molecules. However, current computational methods such as quantum mechanics require a large amount of computation and long computation time, making them unsuitable for large-scale calculations. Deep learning methods also have poor long-range prediction capabilities, and there is a need to improve their accuracy.

[0128] The DeepQC model proposed in the embodiment of this application aims to maintain high accuracy while keeping computational complexity low. To more appropriately verify the effectiveness of the DeepQC model, it was validated on a test set, and its computational accuracy was tested on two datasets, Conformer Benchmark and TorsionNet500, compared with computational chemistry methods, semi-empirical quantum mechanics methods, and deep learning methods. The vertical axis in Figure 6 shows the correlation between molecular energy calculated by various methods in Conformer Benchmark and theoretical values, and the vertical axis in Figure 7 shows the correlation between molecular energy calculated by various methods in TorsionNet500 and theoretical values, where the theoretical values ​​are molecular energy calculated by high-precision quantum mechanics methods. Regarding algorithm speed, when comparing the DeepQC model proposed in the embodiment of this application with high-precision quantum mechanics methods, the DeepQC model shows a speed improvement of several hundred times.

[0129] Next, an exemplary structure of an artificial intelligence-based molecular processing device 455 implemented as a software module according to the embodiments of the present invention will be described. In some embodiments, as shown in Figure 3, the software module in the artificial intelligence-based molecular processing device 455 stored in memory 450 includes an acquisition module 4551 configured to acquire the three-dimensional structure of a target molecule, and a neural network module 4552 configured to call a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule and obtain the energy error of the target molecule, wherein the neural network model is trained by fitting the energy error of a sample molecule, and the energy error of the sample molecule is calculated according to two types of energy calculation mechanisms. The system comprises a neural network module 4552, which includes a first energy calculation process and a second energy calculation process, where the accuracy of the first energy calculation process is less than that of the second energy calculation process, and the speed of the first energy calculation process is faster than that of the second energy calculation process; a calculation module 4553 configured to perform a first energy calculation process on the three-dimensional structure of a target molecule to obtain the first energy of the target molecule; and a correction module 4554 configured to perform an error correction process on the first energy of the target molecule based on the energy error of the target molecule to obtain the second energy of the target molecule.

[0130] In some embodiments, the neural network module 4552 is further configured to perform feature extraction on a three-dimensional structure using a neural network model to obtain energy error features of the target molecule, and then perform fully connected processing on the energy error features using the neural network model to obtain the energy error of the target molecule.

[0131] In some embodiments, the neural network model includes N cascaded feature networks, and the neural network module 4552 further performs initial feature extraction processing for each atom in the three-dimensional structure to obtain the initial features of each atom. If the value of n is 1 ≤ n ≤ N-1, the nth feature network of the cascaded N feature networks is used to perform nth feature extraction processing on the input of the nth feature network to obtain the nth feature of each atom, and the nth feature is transmitted to the (n+1)th feature network. If the value of n is N, attribute feature extraction is performed on the (n-1)th feature of each atom using the nth feature network. The system is configured to perform output processing to obtain the nth attribute feature of each atom, perform coordinate feature extraction processing on the (n-1)th feature of each atom using the nth feature network to obtain the nth coordinate feature of each atom, and then form an energy error feature using the nth attribute feature and the nth coordinate feature of each atom. Here, the range of the value of N satisfies 2 ≤ N, the value of n is an integer increasing from 1, and the range of the value of n satisfies 1 ≤ n ≤ N-1. When the value of n is 1, the input to the nth feature network is the initial feature of each atom, and when the value of n is 2 ≤ n ≤ N, the input to the nth feature network is the (n-1)th feature of each atom output by the (n-1)th feature network.

[0132] In some embodiments, the neural network module 4552 further obtains initial attribute features of each atom in the three-dimensional structure, and initial coordinate features of each atom in the three-dimensional structure, where the initial attribute features represent the attribute information of the atom, and the initial coordinate features represent the position information of the atom. For each atom in the three-dimensional structure, it obtains at least one other atom excluding the atom in the three-dimensional structure, and performs a process to obtain initial relationship features between the atom and each other atom, where the initial relationship features represent the bonding relationship between the atom and the other atoms. The module is configured to form the initial features of each atom using the initial attribute features, initial coordinate features, and initial relationship features of each atom.

[0133] In some embodiments, the neural network module 4552 is further configured to use an nth feature network to perform the following processing for each atom, the processing comprising: obtaining the atoms other than the atoms in the three-dimensional structure excluding the atoms; performing a first mapping process on the (n-1)th feature of an atom and the (n-1)th feature of each other atom to obtain the nth association feature of the atom corresponding to each other atom; and performing a second mapping process on the (n-1)th feature of an atom and the nth association feature of the atom corresponding to each other atom to obtain the nth feature of an atom.

[0134] In some embodiments, the neural network module 4552 is further configured to extract the (n-1)th coordinate feature of an atom from the (n-1)th feature of an atom, extract the (n-1)th coordinate feature of each other atom from the (n-1)th feature of each other atom, extract the (n-1)th attribute feature of an atom from the (n-1)th feature of an atom, extract the (n-1)th attribute feature of each other atom from the (n-1)th feature of each other atom, extract the (n-1)th relation feature of an atom from the (n-1)th feature of an atom, and perform the following processing for each other atom, the processing includes: extracting the (n-1)th relation feature of an atom to another atom from the (n-1)th relation feature of an atom; obtaining a first feature distance between the (n-1)th coordinate feature of an atom and the (n-1)th coordinate feature of another atom; and performing a first fusion process on the square of the first feature distance, the (n-1)th attribute feature of an atom, the (n-1)th attribute feature of the other atom, and the (n-1)th relation feature of an atom to another atom to obtain the nth association feature of an atom corresponding to the other atom.

[0135] In some embodiments, the neural network module 4552 is further configured to perform an addition operation on the nth association features of atoms corresponding to multiple other atoms to obtain the nth association feature of an atom, perform a second fusion operation on the (n-1)th attribute feature and the nth association feature of an atom to obtain the nth attribute feature of an atom, obtain a first feature difference between the (n-1)th coordinate feature of an atom and the (n-1)th coordinate feature of each other atom, perform a weighted average operation on the first feature difference of multiple other atoms using the nth association feature of each other atom as a weight to obtain a weighted average result corresponding to an atom, perform an addition operation on the weighted average result of an atom and the (n-1)th coordinate feature of an atom to obtain the nth coordinate feature of an atom, use the initial relationship feature of an atom as the nth relationship feature of an atom, and form the nth feature of an atom using the nth relationship feature of an atom, the nth attribute feature of an atom, and the nth coordinate feature of an atom.

[0136] In some embodiments, the neural network module 4552 is further configured to extract the (n-1)th coordinate feature of an atom from the (n-1)th feature of an atom, extract the (n-1)th coordinate feature of each other atom from the (n-1)th feature of each other atom, extract the (n-1)th attribute feature of an atom from the (n-1)th feature of an atom, extract the (n-1)th attribute feature of each other atom from the (n-1)th feature of each other atom, extract the (n-1)th relation feature of an atom from the (n-1)th feature of an atom, and perform the following processing for each other atom, the processing being the steps of extracting the (n-1)th relation feature of an atom to other atoms and the (n-1)th coordinate feature of an atom The process includes the steps of obtaining a second feature distance between the (n-1)th coordinate feature of another atom, and performing a first fusion process on the square of the second feature distance, the (n-1)th attribute feature of the atom, the (n-1)th attribute feature of the other atom, and the (n-1)th relation feature of the atom to the other atom to obtain the nth association feature of the atom corresponding to the other atom, wherein if there are multiple other atoms, the process includes the steps of performing an addition process on the nth association features of the atoms corresponding to multiple other atoms to obtain the nth association feature of the atom, and performing a second fusion process on the (n-1)th attribute feature and the nth association feature of the atom to obtain the nth attribute feature of the atom.

[0137] In some embodiments, the device includes a training module 4555 configured to call a neural network model to perform energy error prediction processing on the three-dimensional structure of a target molecule, and before obtaining the energy error of the target molecule, to further acquire a sample molecule, perform a three-dimensional structure generation process on the sample molecule to obtain multiple three-dimensional structures of the sample molecule, obtain the label energy error of each three-dimensional structure of the sample molecule, forward propagate each three-dimensional structure of the sample molecule in the neural network model to obtain the predicted energy error of each three-dimensional structure of the sample molecule, determine the total loss corresponding to the neural network model based on the label energy error of each three-dimensional structure of the sample molecule and the predicted energy error of each three-dimensional structure of the sample molecule, backpropagate the total loss in the neural network model to obtain the parameter change value of the neural network model when the total loss converges, and update the parameters of the neural network model based on the parameter change value.

[0138] In some embodiments, the training module 4555 is further configured to perform the following processing on each sample molecular structure, the processing including the steps of: performing a first energy calculation on the sample molecular structure to obtain a first energy of the sample molecular structure; performing a second energy calculation on the sample molecular structure to obtain a second energy of the sample molecular structure; and obtaining a first difference between the second energy of the sample molecule and the first energy of the sample molecule to be used as the label energy error of the sample molecular structure.

[0139] In some embodiments, the first energy calculation process is an energy calculation process based on semi-empirical quantum mechanics, and the second energy calculation process is an energy calculation process based on density functionals.

[0140] In some examples, training module 4555 further performs a process to determine the root mean square error of the sample molecular structure for each sample molecular structure based on the label energy error of the sample molecular structure and the predicted energy error of the sample molecular structure, obtains the other sample molecular structures excluding the sample molecular structure of the sample molecule, determines the second difference between the label energy error of the sample molecular structure and the label energy error of the other sample molecular structures for each other sample molecular structure, and compares the predicted energy error of the sample molecular structure with the predicted energy error of the other sample molecular structures. The system is configured to perform a process to determine the third difference between the measured energy error, apply the root mean square operation to the second and third differences to obtain the second root mean square error of the sample molecular structure corresponding to other sample molecular structures, perform an addition operation on the second root mean square errors of the sample molecular structures corresponding to multiple other sample molecular structures to obtain the third root mean square error of the sample molecular structure, and perform a third fusion operation on the first root mean square errors of multiple sample molecular structures and the third root mean square errors of multiple sample molecules to obtain the total loss corresponding to the neural network model.

[0141] Embodiments of the present application provide a computer program product that includes a computer program or computer executable instructions stored in a computer-readable storage medium. The processor of an electronic device reads the computer executable instructions from the computer-readable storage medium, and the processor executes the computer executable instructions, thereby causing the electronic device to perform the above-described molecular processing method based on artificial intelligence of the embodiment of the present application.

[0142] The embodiment of the present invention provides a computer-readable storage medium that stores computer-executable instructions for causing a processor to execute a molecular processing method based on artificial intelligence according to the embodiment of the present invention, for example, the molecular processing method based on artificial intelligence shown in Figures 4A to 4C, when executed by the processor.

[0143] In some embodiments, the computer-readable storage medium may be memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, surface-mount memory, optical disk, or CD-ROM, or it may be a variety of devices that include one of the above memories or any combination thereof.

[0144] In some embodiments, computer executable instructions may take the form of programs, software, software modules, scripts, or code, and may be written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be arranged in any form, such as being a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0145] For example, computer executable instructions may, but not necessarily, correspond to files in a file system; they can be stored in parts of files that store other programs or data, for example, in one or more scripts within a Hyper Text Markup Language (HTML) document, in a single file of the program discussed, or in multiple collaborative files (e.g., files that store one or more modules, subprograms, or code portions).

[0146] For example, a computer executable instruction may be arranged to run on one electronic device, or on multiple electronic devices located in the same place, or on multiple electronic devices distributed in multiple locations and interconnected by a communication network.

[0147] In summary, the first energy of the target molecule is obtained by performing a first energy calculation process on the three-dimensional structure of the target molecule in the embodiment of this application. Since the speed of the first energy calculation process is faster than the second energy calculation process, the energy calculation process is faster, and an energy error prediction process is performed on the three-dimensional structure of the target molecule using a neural network model to obtain the energy error of the target molecule. An error correction process is then performed on the first energy calculated based on the energy error to obtain the second energy of the target molecule. Since the energy error can represent the difference in calculation results between the highly accurate second energy calculation process and the less accurate first energy calculation process, the accuracy of the second energy can be improved by correcting the first energy calculated using the energy error predicted by deep learning.

[0148] The above description is merely an example of the present application and does not limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made in the spirit and scope of this application are all included within the scope of protection of this application.

Claims

1. An artificial intelligence-based molecular processing method performed by an electronic device, Steps to obtain the three-dimensional structure of the target molecule, A step of calling a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule and obtaining the energy error of the target molecule, wherein the neural network model is trained by inputting the sample molecule's three-dimensional structure and fitting the energy error of the sample molecule obtained therefrom, wherein the energy error of the sample molecule refers to the difference in calculation results obtained when calculating the energy of the sample molecule according to two types of energy calculation mechanisms, the two types of energy calculation mechanisms include a first energy calculation process and a second energy calculation process, the accuracy of the first energy calculation process is lower than the accuracy of the second energy calculation process, and the speed of the first energy calculation process is faster than the speed of the second energy calculation process, The steps include performing the first energy calculation process on the three-dimensional structure of the target molecule to obtain the first energy of the target molecule, A molecular processing method based on artificial intelligence, comprising the steps of: performing an error correction process on the first energy of the target molecule based on the energy error of the target molecule to obtain the second energy of the target molecule.

2. The step of calling the neural network model, performing energy error prediction processing on the three-dimensional structure of the target molecule, and obtaining the energy error of the target molecule is: The steps include: performing a feature extraction process on the three-dimensional structure using the neural network model to obtain the energy error features of the target molecule; The molecular processing method according to claim 1, comprising the step of performing a fully connected process on the energy error features using the neural network model to obtain the energy error of the target molecule.

3. The neural network model includes N cascaded feature networks, and the step of using the neural network model to perform feature extraction on the three-dimensional structure to obtain the energy error features of the target molecule is as follows: The steps include: performing an initial feature extraction process on each atom within the three-dimensional structure to obtain the initial features of each atom; If the value of n is 1 ≤ n ≤ N-1, the nth feature network among the n cascaded feature networks is used to perform an nth feature extraction process on the input of the nth feature network to obtain the nth feature of each atom, and the nth feature is transmitted to the (n+1)th feature network. If the value of n is N, the process includes the steps of: performing attribute feature extraction processing on the (n-1)th feature of each atom using the n-feature network to obtain the nth attribute feature of each atom; performing coordinate feature extraction processing on the (n-1)th feature of each atom using the n-feature network to obtain the nth coordinate feature of each atom; and forming the energy error feature using the nth attribute feature and the nth coordinate feature of each atom. The molecular processing method according to claim 2, wherein the range of the value of N satisfies 2 ≤ N, the value of n is an integer increasing from 1, and the range of the value of n satisfies 1 ≤ n ≤ N-1, and when the value of n is 1, the input to the n-feature network is the initial feature of each atom, and when the value of n is 2 ≤ n ≤ N, the input to the n-feature network is the (n-1) feature of each atom output by the (n-1)-feature network.

4. The step of performing an initial feature extraction process on each atom within the three-dimensional structure to obtain the initial features of each atom is: A step of obtaining initial attribute characteristics of each atom in the three-dimensional structure and obtaining initial coordinate characteristics of each atom in the three-dimensional structure, wherein the initial attribute characteristics represent attribute information of the atom and the initial coordinate characteristics represent position information of the atom. A step of obtaining, for each of the atoms in the three-dimensional structure, at least one other atom excluding the atom in the three-dimensional structure, and performing a process to obtain initial relationship characteristics between the atom and each of the other atoms, wherein the initial relationship characteristics represent the bonding relationship between the atom and the other atoms. A molecular processing method according to claim 3, comprising the step of forming the initial characteristics of each atom using the initial attribute characteristics of each atom, the initial coordinate characteristics of each atom, and the initial relational characteristics of each atom.

5. The step of using the nth feature network among the cascaded N feature networks to perform an nth feature extraction process on the input of the nth feature network to obtain the nth feature of each atom is as follows: The process includes the step of performing a process on each of the atoms using the n feature network, the process being: A step of obtaining other atoms from the three-dimensional structure, excluding the aforementioned atoms. A step of performing a first mapping process on the (n-1)th characteristic of the atom and the (n-1)th characteristic of each of the other atoms to obtain the nth association characteristic of the atom corresponding to each of the other atoms, A molecular processing method according to claim 3, comprising the step of performing a second mapping process on the (n-1)th characteristic of the atom and the nth association characteristic of the atom corresponding to each of the other atoms to obtain the nth characteristic of the atom.

6. The step of performing a first mapping process on the (n-th)th feature of the atom and the (n-th)th feature of each of the other atoms to obtain the nth association feature of the atom corresponding to each of the other atoms is: The steps include extracting the (n-th) coordinate feature of the atom, the (n-th) attribute feature of the atom, and the (n-th) relation feature of the atom from the (n-th) feature of the atom, The steps include extracting the (n-th)th coordinate feature and the (n-th)th attribute feature of each of the other atoms from the (n-th)th feature of each of the other atoms, The process includes, for each of the other atoms, performing a process, The steps include extracting the n-th-th relational feature of the atom with respect to the other atoms from the n-th-th relational feature of the atom, A step of obtaining a first feature distance between the (n-1) coordinate feature of the atom and the (n-1) coordinate feature of the other atom, A molecular processing method according to claim 5, comprising the step of performing a first fusion process on the square of the first feature distance, the (n-1)th attribute feature of the atom, the (n-1)th attribute feature of the other atom, and the (n-1)th relation feature of the atom with respect to the other atom to obtain the nth relation feature of the atom corresponding to the other atom.

7. The step of performing a second mapping process on the (n-1)th characteristic of the atom and the nth association characteristic of the atom corresponding to each of the other atoms to obtain the nth characteristic of the atom is: A step of performing an addition process on the nth association feature of the atom corresponding to multiple other atoms to obtain the nth association feature of the atom, A step of performing a second fusion process on the (n-1) attribute feature of the atom and the nth association feature of the atom to obtain the nth attribute feature of the atom, A step of obtaining a first feature difference between the (n-1) coordinate feature of the atom and the (n-1) coordinate feature of each of the other atoms, A step of performing a linear mapping process on the nth association feature of the atom corresponding to each of the other atoms to obtain the weight of each of the other atoms, A step of performing a weighted average operation on the first feature difference of a plurality of the other atoms based on the weight of each of the other atoms, and obtaining a weighted average result corresponding to the atom, The steps include performing an addition process on the weighted average result of the atoms and the (n-1)th coordinate feature of the atoms to obtain the nth coordinate feature of the atoms, A molecular processing method according to claim 6, comprising the step of defining the initial relational features of the atom as the nth relational features of the atom, and forming the nth feature of the atom using the nth relational features of the atom, the nth attribute features of the atom, and the nth coordinate features of the atom.

8. The step of performing attribute feature extraction processing on the (n-1)th feature of each atom using the n-feature network to obtain the nth attribute feature of each atom is: The steps include extracting the (n-th)th coordinate feature of the atom, the (n-th)th attribute feature of the atom, and the (n-th)th relation feature of the atom from the (n-th)th feature of each atom, The process includes the steps of obtaining, for each of the aforementioned atoms, other atoms in the three-dimensional structure excluding the aforementioned atom, and performing a process on each of the other atoms, wherein the process is The process involves extracting the (n-1) relational features of the atom with respect to the other atom from the (n-1) relational features of the atom, and obtaining a second feature distance between the (n-1) coordinate features of the atom and the (n-1) coordinate features of the other atom. A step of performing a first fusion process on the square of the second feature distance, the (n-1)th attribute feature of the atom, the (n-1)th attribute feature of the other atom, and the (n-1)th relation feature of the atom with respect to the other atom to obtain the nth association feature of the atom corresponding to the other atom, A step of performing an addition process on the nth association feature of the atom corresponding to multiple other atoms to obtain the nth association feature of the atom, A molecular processing method according to claim 3, comprising the step of performing a second fusion process on the (n-1) attribute feature of the atom and the nth association feature of the atom to obtain the nth attribute feature of the atom.

9. The molecular processing method invokes a neural network model to perform energy error prediction processing on the three-dimensional structure of the target molecule, and before obtaining the energy error of the target molecule, The steps include obtaining a sample molecule, performing a three-dimensional structure generation process on the sample molecule, and obtaining the three-dimensional structures of multiple sample molecules, The steps include obtaining the label energy error for each of the sample molecular structures, The steps include: forward propagation of each sample molecular structure in an initialized neural network model to obtain the predicted energy error for each sample molecular structure; The steps include determining the total loss based on the label energy error of each sample molecular structure and the predicted energy error of each sample molecular structure, A molecular processing method according to any one of claims 1 to 8, further comprising the steps of: performing a backpropagation operation on the total loss using the initialized neural network model to obtain parameter change values ​​of the initialized neural network model when the total loss converges; and updating the parameters of the initialized neural network model based on the parameter change values.

10. The step of obtaining the label energy error for each of the sample molecular three-dimensional structures is: The process includes the step of performing a treatment on each of the sample molecular three-dimensional structures, and the treatment is as follows: The steps include: performing a first energy calculation process on the sample molecular three-dimensional structure to obtain the first energy of the sample molecular three-dimensional structure; The steps include performing a second energy calculation process on the sample molecular three-dimensional structure to obtain the second energy of the sample molecular three-dimensional structure, The molecular processing method according to claim 9, comprising the step of obtaining a first difference between the second energy of the sample molecular three-dimensional structure and the first energy of the sample molecular three-dimensional structure, and using this as the label energy error of the sample molecular three-dimensional structure.

11. The step of determining the total loss based on the label energy error of each sample molecular structure and the predicted energy error of each sample molecular structure is: Regarding the three-dimensional structure of each sample molecule, The steps include: performing a process to determine the root mean square error of the sample molecular structure based on the label energy error of the sample molecular structure and the predicted energy error of the sample molecular structure; The process includes the step of obtaining the three-dimensional structures of other sample molecules excluding the three-dimensional structure of the sample molecule, The steps include: determining a second difference between the label energy error of the sample molecular structure and the label energy error of the other sample molecular structure for each of the other sample molecular structures, and performing a process to determine a third difference between the predicted energy error of the sample molecular structure and the predicted energy error of the other sample molecular structure; The steps include: performing root mean square processing on the second difference and the third difference to obtain the second root mean square error of the sample molecular structure corresponding to the other sample molecular structure; The steps include: performing an addition process on the root mean square errors of the sample molecular three-dimensional structures corresponding to multiple other sample molecular three-dimensional structures to obtain the root mean square error of the sample molecular three-dimensional structure; A molecular processing method according to claim 9, comprising the step of performing a processing that includes: performing a third fusion process on the first root mean square error and the third root mean square error of the multiple sample molecular three-dimensional structures to obtain a total loss corresponding to the neural network model.

12. A molecular processing device based on artificial intelligence, An acquisition module configured to acquire the three-dimensional structure of a target molecule, A neural network module configured to invoke a neural network model, perform energy error prediction processing on the three-dimensional structure of the target molecule, and obtain the energy error of the target molecule, The neural network model is trained by inputting the three-dimensional structure of a sample molecule and fitting the energy error of the sample molecule, wherein the energy error of the sample molecule refers to the difference in calculation results obtained when calculating the energy of the sample molecule according to two types of energy calculation mechanisms, the two types of energy calculation mechanisms include a first energy calculation process and a second energy calculation process, the accuracy of the first energy calculation process is lower than the accuracy of the second energy calculation process, and the speed of the first energy calculation process is faster than the speed of the second energy calculation process, the neural network module, A computation module configured to perform the first energy calculation process on the three-dimensional structure of the target molecule and obtain the first energy of the target molecule, An artificial intelligence-based molecular processing apparatus comprising: a correction module configured to perform error correction processing on the first energy of the target molecule based on the energy error of the target molecule, thereby obtaining the second energy of the target molecule.

13. Memory that stores executable computer instructions, An electronic device comprising: a processor that executes computer-executable instructions stored in the memory to perform the molecular processing method based on artificial intelligence as described in any one of claims 1 to 8.

14. A computer program that causes a computer to execute a molecular processing method based on artificial intelligence as described in any one of claims 1 to 8.

Images