A method, apparatus, medium, and program product for constructing a machine learning force field for non-equilibrium biological system simulations

Abstract

The purpose of the present application is to provide a method, device, medium and program product for constructing a machine learning force field for non-equilibrium biological system simulation, the method comprising: constructing a quantum mechanics simulation model corresponding to a biological system; based on the quantum mechanics simulation model, obtaining a corresponding non-equilibrium data set through transition state scanning, wherein the non-equilibrium data set comprises structure information and corresponding system state information of the biological system at each reaction coordinate; and based on the non-equilibrium data set and an equilibrium data set, training a corresponding machine learning force field. By obtaining non-equilibrium data for machine learning force field training, high-precision energy prediction can be achieved at non-standard bond lengths (such as transition states of chemical reactions), so that molecular dynamics simulation of non-equilibrium systems (such as biological system chemical reactions) can be performed using the machine learning force field, and high-precision biochemical property prediction can be performed.

Description

Technical Field

[0001] This application relates to the field of bioinformatics technology, and in particular to a technique for constructing machine learning force fields for simulating non-equilibrium biological systems. Background Technology

[0002] Biological systems are dynamic systems composed of biomolecules (e.g., proteins, nucleic acids, small molecules) and their interactions. Research on biological systems helps us understand the interaction mechanisms of these molecules, providing a theoretical basis for drug development and disease treatment.

[0003] Most machine learning force fields (MLFFs) used for simulating biological systems are trained on equilibrium or near-equilibrium conformations sampled from classical force field simulations. However, when these MLFFs are applied to systems outside the training set distribution, such as predicting transition states of chemical reactions, their accuracy can drop significantly. Summary of the Invention

[0004] One object of this application is to provide a method, apparatus, medium, and program product for constructing machine learning force fields for simulating non-equilibrium biological systems.

[0005] According to one aspect of this application, a method for constructing a machine learning force field for simulating non-equilibrium biological systems is provided, the method comprising:

[0006] Construct a quantum mechanical simulation model corresponding to a biological system, wherein the quantum mechanical simulation model includes a quantum mechanical cluster model or a combined model of quantum mechanics and molecular dynamics;

[0007] Based on the quantum mechanical simulation model, a corresponding non-equilibrium state dataset is obtained through transition state scanning. The non-equilibrium state dataset includes the structural information of the biological system and the corresponding system state information under each reaction coordinate.

[0008] Based on the non-equilibrium dataset and the equilibrium dataset, a corresponding machine learning force field is trained and obtained. The equilibrium dataset includes the structural information of the equilibrium conformation corresponding to the biological system and the corresponding system state information.

[0009] According to one aspect of this application, a computer device for constructing a machine learning force field for simulating non-equilibrium biological systems is provided, comprising a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of any of the methods described above.

[0010] According to one aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0011] According to one aspect of this application, a computer program product is provided, comprising a computer program, characterized in that, when executed by a processor, the computer program implements the steps of any of the methods described above.

[0012] According to one aspect of this application, an apparatus is provided for constructing a machine learning force field for simulating non-equilibrium biological systems, the apparatus comprising:

[0013] A module for constructing quantum mechanical simulation models corresponding to biological systems, wherein the quantum mechanical simulation model includes a quantum mechanical cluster model or a combined model of quantum mechanics and molecular dynamics;

[0014] The first and second modules are used to obtain the corresponding non-equilibrium state dataset by means of transition state scanning based on the quantum mechanical simulation model. The non-equilibrium state dataset includes the structural information of the biological system and the corresponding system state information under each reaction coordinate.

[0015] The first and third modules are used to train and obtain the corresponding machine learning force field based on the non-equilibrium dataset and the equilibrium dataset, wherein the equilibrium dataset includes the structural information of the equilibrium conformation corresponding to the biological system and the corresponding system state information.

[0016] Compared with existing technologies, this application constructs a quantum mechanical simulation model corresponding to a biological system. This quantum mechanical simulation model includes a quantum mechanical cluster model or a combined model of quantum mechanics and molecular dynamics. Based on the quantum mechanical simulation model, a corresponding non-equilibrium state dataset is obtained through transition state scanning. This non-equilibrium state dataset includes the structural information and corresponding system state information of the biological system at various reaction coordinates. Based on the non-equilibrium state dataset and the equilibrium state dataset, a corresponding machine learning force field is trained. The equilibrium state dataset includes the structural information and corresponding system state information of the equilibrium conformation of the biological system. By acquiring non-equilibrium state data for machine learning force field training, high-precision energy prediction can be achieved for non-standard bond lengths (e.g., transition states of chemical reactions). Therefore, this machine learning force field can be used to perform molecular dynamics simulations of non-equilibrium systems (e.g., chemical reactions in biological systems) and to predict biochemical properties with high accuracy. Attached Figure Description

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0018] Figure 1 A flowchart illustrating a method for constructing a machine learning force field for simulating a non-equilibrium biological system according to an embodiment of this application is shown.

[0019] Figure 2 A schematic diagram of a device structure for constructing a machine learning force field for simulating non-equilibrium biological systems, according to an embodiment of this application, is shown.

[0020] Figure 3 Exemplary systems that can be used to implement the various embodiments described in this application are shown.

[0021] The same or similar reference numerals in the accompanying drawings represent the same or similar parts. Detailed Implementation

[0022] The present application will now be described in further detail with reference to the accompanying drawings.

[0023] In a typical configuration of this application, the terminal, the device of the service network, and the trusted party all include one or more processors (e.g., a central processing unit (CPU)), input / output interfaces, network interfaces, and memory.

[0024] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory. Memory is an example of computer-readable media.

[0025] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PCM), programmable random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0026] The devices referred to in this application include, but are not limited to, user equipment, network equipment, or devices composed of user equipment and network equipment integrated through a network. The user equipment includes, but is not limited to, any mobile electronic product capable of human-computer interaction (e.g., via a touchpad), such as smartphones and tablets. These mobile electronic products can use any operating system, such as Android or iOS. The network equipment includes an electronic device capable of automatically performing numerical calculations and information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), and embedded devices. The network equipment includes, but is not limited to, computers, network hosts, single network servers, multiple network server clusters, or clouds composed of multiple servers. Here, a cloud consists of a large number of computers or network servers based on cloud computing, where cloud computing is a type of distributed computing, consisting of a virtual supercomputer composed of a group of loosely coupled computer clusters. The network includes, but is not limited to, the Internet, wide area network, metropolitan area network, local area network, VPN network, wireless ad hoc network, etc. Preferably, the device can also be a program running on the user equipment, network device, or a device formed by integrating user equipment and network device, network device, touch terminal, or network device and touch terminal through a network.

[0027] Of course, those skilled in the art should understand that the above-described devices are merely examples, and other existing or future devices that are applicable to this application should also be included within the scope of protection of this application, and are hereby incorporated by reference.

[0028] In the description of this application, "multiple" means two or more, unless otherwise expressly and specifically defined.

[0029] Figure 1A flowchart illustrating a method for constructing a machine learning force field for simulating a non-equilibrium biological system according to an embodiment of this application is shown. The method includes steps S11, S12, and S13. In step S11, device 1 constructs a quantum mechanical simulation model corresponding to the biological system, wherein the quantum mechanical simulation model includes a quantum mechanical cluster model or a combined model of quantum mechanics and molecular dynamics. In step S12, based on the quantum mechanical simulation model, device 1 obtains a corresponding non-equilibrium dataset through transition state scanning, wherein the non-equilibrium dataset includes the structural information of the biological system at each reaction coordinate and the corresponding system state information. In step S13, device 1 trains and obtains a corresponding machine learning force field based on the non-equilibrium dataset and the equilibrium dataset, wherein the equilibrium dataset includes the structural information of the equilibrium conformation corresponding to the biological system and the corresponding system state information.

[0030] In some embodiments, the device 1 includes, but is not limited to, user equipment or network devices with information processing or computing capabilities, such as tablet computers, computers, and servers. The machine learning force field is constructed within the device 1.

[0031] In step S11, device 1 constructs a quantum mechanical simulation model corresponding to the biological system, wherein the quantum mechanical simulation model includes a quantum mechanical cluster model or a combined model of quantum mechanics and molecular dynamics.

[0032] The construction of the quantum mechanical cluster (QM Cluster) model or the quantum mechanical and molecular dynamics (QM / MM) combined model typically starts from the three-dimensional structure of the biological system. For the quantum mechanical cluster model, the structural information of biomolecules in the biological system (e.g., PDB files) is input into molecular visualization software (e.g., PyMol, Chimera, or ChimeraX), and the active sites are extracted using this software. The active sites are typically selected from the substrate and atoms within a predetermined distance, and the number of atoms extracted from the active sites is usually ≤300. For the quantum mechanical and molecular dynamics combined model, the structural information of biomolecules in the biological system (e.g., PDB files) is input into molecular visualization software (e.g., PyMol (https: / / pymol.org / ), Chimera (https: / / www.cgl.ucsf.edu / chimera / ), or ChimeraX (https: / / www.cgl.ucsf.edu / chimerax / ), and the QM and MM regions are distinguished using this software. The QM region is typically selected from the substrate and atoms within a predetermined distance, with the number of atoms extracted from the active center usually ≤300. The MM region includes the remaining protein structure, solvent, ions, etc., after the QM region is defined. The aforementioned active center or the structure with the QM and MM regions separated is imported into GaussView (https: / / gaussian.com / ) for processing to obtain the corresponding quantum mechanical cluster model or a combined quantum mechanical and molecular dynamics model. For example, by using GaussView to perform hydrogenation or chemical bond correction on the aforementioned active center or QM region structure, the resulting structure serves as the quantum mechanical simulation model for QM calculations in subsequent transition state scans.

[0033] In step S12, device 1 obtains the corresponding non-equilibrium state dataset by means of transition state scanning based on the quantum mechanical simulation model. The non-equilibrium state dataset includes the structural information of the biological system and the corresponding system state information under each reaction coordinate.

[0034] In some embodiments, the structural information of the biological system at each reaction coordinate includes the atomic coordinates, atomic number, total charge, and spin state information of the corresponding atoms in the biological system at the corresponding reaction coordinates. The spin state information can be represented by the number of unpaired electrons. The number of unpaired electrons is determined based on the structural information, combined with the corresponding electron configuration rules (e.g., Hund's rules, Pauli exclusion principle, etc.). The system state information of the biological system at each reaction coordinate includes the total energy information of the biological system at the corresponding reaction coordinate and the force information of the corresponding atoms. In some embodiments, the Nudged Elastic Band (NEB) algorithm or the Relaxed Scan algorithm can be used for transition state scanning. The transition state scanning process includes: determining the reaction coordinate (RC) and corresponding scanning information (e.g., scanning range and scanning step size) based on the reaction mechanism; gradually changing the molecular conformation of the biological system based on the reaction coordinate and the corresponding scanning information, determining the atomic coordinates, force information, and total energy corresponding to each step, and outputting the scanning data. The reaction coordinates are variables used to describe the reaction process between biomolecules in the biological system. These coordinates include, but are not limited to, bond lengths, bond angles, or dihedral angles. The reaction mechanism describes the microscopic processes of biomolecular interactions in the biological system. Based on this mechanism, the reaction process can be understood, and appropriate reaction coordinates and scanning information can be set. For example, in a nucleophilic attack reaction, the reaction coordinates can be set as the distance d between the nucleophilic atom and the electrophilic center (usually a carbon atom). C-Nu Based on this distance d C-Nu The changes throughout the reaction process are monitored, and the corresponding scan range is set as follows. Scan step size Based on a quantum mechanical simulation model, the distance d is gradually reduced within the scanning range according to a set step size. C-Nu Quantum mechanical calculations are used to calculate the atomic coordinates, force information, and total energy information corresponding to the molecular conformation in the reaction coordinates for each step, and then the corresponding non-equilibrium state dataset is obtained based on this information. In some embodiments, quantum chemistry software such as Gaussian (https: / / gaussian.com / ), ORCA (https: / / www.faccts.de / orca / ), or PySCF (https: / / pyscf.org / ) can be used to perform the quantum mechanical calculations in the aforementioned transition state scan.

[0035] In some embodiments, step S12 includes: performing a transition state scan based on the quantum mechanical simulation model to obtain corresponding scan data, wherein the scan data includes atomic coordinate information, atomic force information and total energy information of the biological system corresponding to each reaction coordinate during the transition state scan; and preprocessing the scan data to obtain the non-equilibrium state dataset.

[0036] In some embodiments, the scanning data is acquired through the aforementioned transition state scanning process. The scanning data includes atomic coordinates, atomic force information, and total energy information of the molecular conformations corresponding to the biological system at each reaction coordinate throughout the entire reaction process. That is, the scanning data includes a large amount of non-equilibrium atomic coordinates, atomic force information, and total energy information. This scanning data can be used as a non-equilibrium dataset. In processing the scanning data, if the quantum mechanical simulation model used is a combined quantum mechanics and molecular dynamics model, then when constructing the non-equilibrium dataset, only the atomic coordinates, atomic force information, and total energy information corresponding to the QM region are read; it is not necessary to read the information in the MM region, thereby ensuring that the non-equilibrium dataset contains high-precision information. In some embodiments, since the scanning data includes data at different reaction coordinates throughout the entire reaction process, it also contains a very small amount of equilibrium state data. In processing the scanning data, the data of equilibrium conformations can be removed. Of course, since the proportion of equilibrium state data in the scanning data is extremely low, this part of the data may not be removed to improve processing efficiency. In some embodiments, to facilitate subsequent data reading and machine learning force field training, the scanned data can be converted into a preset format during preprocessing and saved as a corresponding non-equilibrium dataset. For example, the scanned data can be converted into the torch.Tensor format in PyTorch and stored as the InMemoryDataset format in PyTorch Geometric. This allows for quick loading into the GPU for machine learning force field training during training, and the use of PyTorch's DataLoader and other related functions (e.g., training with multiple datasets, and extracting information from different datasets according to set probabilities during training).

[0037] In step S13, device 1 trains and obtains a corresponding machine learning force field based on the non-equilibrium dataset and the equilibrium dataset, wherein the equilibrium dataset includes the structural information of the equilibrium conformation corresponding to the biological system and the corresponding system state information.

[0038] In some embodiments, the structural information of the equilibrium conformation corresponding to the biological system includes the atomic coordinates, atomic number, total charge, and spin state information under the equilibrium conformation. The system state information of the equilibrium conformation corresponding to the biological system includes the total energy information, the force information of the corresponding atoms, and the dipole moment information under the equilibrium conformation. The equilibrium state dataset can be obtained from open-source datasets or generated independently.

[0039] In some embodiments, the method further includes: step S14 (not shown), where device 1 performs molecular dynamics simulation based on the biological system to obtain multiple corresponding frame information; uses a molecular cutting algorithm to determine multiple structural information corresponding to each frame information, and uses quantum chemical methods to calculate the system state information corresponding to each structural information; and determines the corresponding equilibrium state dataset based on the structural information and the corresponding system state information.

[0040] For example, through long-term molecular dynamics simulations, the biological system can traverse its phase space and reach equilibrium, allowing for sufficient sampling and acquisition of multiple frame information. Each frame information represents a snapshot of the biological system at a specific time, including but not limited to the atomic number, total charge, and atomic coordinates of each atom in the biological system at the corresponding time. To improve sampling efficiency, enhanced sampling methods such as metadynamics or umbrella sampling can be used to address high energy barriers and long time scales in the simulation, thereby obtaining frame information at equilibrium.

[0041] For each frame of information, several atoms are randomly selected as atoms of interest from the corresponding biological system. For each atom of interest, a molecular cutting algorithm is used to determine a candidate structure for that frame. For example, a sphere is cut with the atom of interest as the center and a preset distance threshold as the radius. A corresponding atom set is established based on all atoms within the sphere. It is determined whether the boundary atoms in the atom set meet the preset cutting conditions. The boundary atoms have chemical bonds that are cut, meaning that atoms connected to the boundary atoms through these cut chemical bonds do not belong to the atom set. The preset cutting conditions include: the chemical bonds cut by the boundary atoms are single bonds; and the boundary atoms are not hydrogen atoms. If all boundary atoms meet the preset cutting conditions, the boundary atoms are hydrogenated based on the chemical bonds that are cut; the candidate structure for the biological system after hydrogenation is determined. If any boundary atoms do not meet the preset cutting conditions, atoms connected to the boundary atoms that do not meet the conditions through these cut chemical bonds are added to the atom set, thereby updating the atom set. The determination of the preset cutting conditions continues. For each candidate structural information, appropriate analytical tools can be used to eliminate those with anomalies (e.g., chemical bond lengths exceeding the appropriate range, abnormal atomic connection numbers, or non-closed ring structures), thus determining multiple structural information. Quantum chemical methods (perturbation theory, density functional theory, etc.) are then used to calculate the system state information corresponding to each structural information. Finally, each structural information and its corresponding system state information are compiled into an equilibrium state dataset.

[0042] In some embodiments, the training architecture of the machine learning force field is a deep neural network. The deep neural network includes an input layer, hidden layers, and an output layer. The hidden layer includes an embedding layer and an interaction module layer. The embedding layer receives structural information from the input layer and converts the input structural information into a corresponding structural information embedding representation. The interaction module layer includes multiple interaction modules, including local interaction modules and non-local interaction modules. The local interaction modules update the structural information based on the structural information embedding representations corresponding to the atoms and their surrounding neighboring atoms, combined with corresponding basis functions, to capture short-range interactions between atoms. The non-local interaction modules use a self-attention mechanism to obtain the interactions between these atoms based on the similarity or correlation of the structural information embedding representations, to capture long-range interactions between atoms. Residual connections are used between the hidden layers. The output layer predicts the corresponding total energy information, the force information of each atom, and / or dipole moment information based on the structural information output by the hidden layers.

[0043] In some embodiments, step S13 includes: step S131 (not shown), where device 1 determines a corresponding training dataset based on the non-equilibrium dataset and the equilibrium dataset, wherein the data from the non-equilibrium dataset and the data from the equilibrium dataset in the training dataset satisfy a preset quantity ratio, and the training dataset includes corresponding structural information and corresponding system state information; step S132 (not shown), where device 1 trains and obtains a corresponding machine learning force field based on the training dataset.

[0044] For example, a predetermined number of data points can be extracted from both the non-equilibrium and equilibrium datasets to train the machine learning force field. Before training with the training dataset, a predetermined number of data points (e.g., 10% of the training dataset) can be reserved as a test set for testing after the machine learning force field training is completed.

[0045] In some embodiments, step S13 further includes: step S133 (not shown), where device 1 determines other training datasets based on the non-equilibrium dataset and the equilibrium dataset, with other preset quantity ratios; trains and obtains corresponding machine learning force fields based on the other training datasets; and selects the final machine learning force field from the machine learning force fields obtained by training the training datasets with each preset quantity ratio.

[0046] For example, machine learning force fields can be trained using training datasets with varying ratios of non-equilibrium to equilibrium data. Before training, a predetermined number of data points can be retained as a test set for testing the trained force fields. For machine learning force fields trained with different predetermined ratios, the one that performs better than others can be selected as the final machine learning force field. The trained machine learning force field can be used to determine its prediction accuracy on both non-equilibrium and equilibrium data using data from the test set. From these machine learning force fields, the one that demonstrates better prediction accuracy on non-equilibrium data than other machine learning force fields, and whose prediction accuracy on equilibrium data is no worse than that of existing machine learning force fields used for equilibrium configurations, can be selected.

[0047] In some embodiments, step S132 includes: device 1 obtaining corresponding system state prediction information based on the centralized structural information in the training dataset using a machine learning force field; determining a corresponding composite loss based on the system state prediction information and the system state information corresponding to the centralized structural information in the training dataset, combined with a composite loss function matching the system state information; and updating the machine learning force field based on the composite loss and a corresponding optimization algorithm.

[0048] For example, machine learning force fields can be used to obtain system state prediction information corresponding to structural information. If the structural information comes from a non-equilibrium dataset, the system state prediction information includes total energy prediction information and force prediction information for each atom. The corresponding composite loss function includes the energy loss function LE and the force loss function LF. If the structural information comes from an equilibrium dataset, the system state prediction information includes total energy prediction information, force prediction information for each atom, and dipole moment prediction information. The corresponding composite loss function includes the energy loss function LE, the force loss function LF, and the dipole moment loss function Lμ. The energy loss function LE is used to measure the difference between the total energy prediction information and the total energy information. The force loss function LF is used to measure the difference between the force prediction information and the force information. The dipole moment loss function Lμ is used to measure the difference between the dipole moment prediction information and the dipole moment information. In some embodiments, the energy loss function LE, the force loss function LF, and the dipole moment loss function Lμ can all be calculated using the mean squared error (MSE) loss function. The corresponding composite loss function L = αELE + αFLF(+αμLμ) is obtained through weighted combination, where αE, αF, and αμ are the corresponding weight hyperparameters. The machine learning force field is updated by minimizing the composite loss using a corresponding optimization algorithm. The optimization algorithm includes, but is not limited to, the Adam algorithm or its variants (e.g., AMSGrad, AdamW, etc.), Newton's method, or quasi-Newton methods.

[0049] In some embodiments, the method further includes: step S15 (not shown), where the device 1 performs molecular dynamics simulations using the machine learning force field based on the enhanced sampling method to determine reaction-related information of the chemical reaction corresponding to the biological system.

[0050] Because the machine learning force field obtained through this training has good predictive ability for transition state systems, it can be deployed in relevant molecular dynamics simulation tools (e.g., ASE (https: / / ase-lib.org / ), SchnetPack (https: / / schnetpack.readthedocs.io / en / latest / ), or OpenMM (https: / / openmm.org / )) to accurately simulate chemical reactions (bond formation and breaking) in biological systems. Furthermore, for events that are difficult to sample due to high reaction energy barriers, enhanced sampling methods will be applied to the simulation to determine the corresponding reaction-related information.

[0051] The molecular dynamics simulation refers to the process of obtaining the trajectories of each atom in the biological system using the basic principles of Newtonian mechanics and by integrating equations of motion; that is, the process of how the coordinates and velocities of each atom in the biological system change over time. For the i-th atom:

[0052]

[0053] Where, m i Let r be the mass of the i-th atom. i Let f be the coordinate of the i-th atom. i Let U be the force acting on the i-th atom, and U be the total energy of the biological system. In the molecular dynamics simulation, the machine learning force field is combined with the velocity Verlet integrator to calculate the forces and total energy and update the coordinates and velocity changes of the atoms.

[0054] Here, the molecular dynamics simulation based on the machine learning force field obtained from this training, as described in this scheme, supports simulations such as NVE (isolated system), NVT (isothermal), and NPT (isothermal and isobaric).

[0055] In some embodiments, the reaction-related information includes reaction rate constant information; step S15 includes: device 1 applies a first bias potential energy matching each force constant value to the reaction coordinates corresponding to the biological system based on different force constant values, performs molecular dynamics simulation using the machine learning force field, and determines the reaction time corresponding to the biological system under each force constant value; based on transition state theory, linearly fits the force constants to each force constant value using the reaction time corresponding to the biological system under each force constant value, and determines the reaction rate constant information.

[0056] In some embodiments, the applied first bias potential energy is U(x) = k·(xx) off ) 2Where k is the force constant, x is the reaction coordinate, and x off The coordinates used to represent the position of the reaction after a reaction occurs (the breaking of old chemical bonds and the formation of new chemical bonds) (i.e., the state of the products) are constants. For example, in the reaction AX......B→A......XB, where A, B, and X represent atoms, - represents covalent bonds, and ... represents non-covalent bonds, in the aforementioned formula for the first bias potential energy, x can be determined as the distance from atom A to X during the reaction. off This is the distance from X to A after X has completely left A.

[0057] Based on the aforementioned formula for the first bias potential energy, the set force constant value k is... i The corresponding first bias potential energy U i (x) is applied to the reaction coordinates, and the aforementioned molecular dynamics simulation is performed in conjunction with the machine learning force field to obtain the trajectory of each atom in the biological system, that is, the process of the coordinates and velocities of each atom in the biological system changing over time. Based on this trajectory, it is detected whether the distance between target atoms in the reaction meets a preset condition, and the time corresponding to the first detection that the preset condition is met is determined as the reaction time. The preset condition includes that the distance between atoms corresponding to old chemical bonds (e.g., the distance from atom A to X) is greater than a first preset threshold (i.e., old chemical bonds break), and / or the distance between atoms corresponding to new chemical bonds (e.g., the distance from atom B to X) is less than a second preset threshold (i.e., new chemical bonds are formed).

[0058] In some embodiments, to more accurately determine the reaction time, the force constant is set to a value k for each setting. i A corresponding first bias potential U can be applied. i (x) Multiple molecular dynamics simulations were run independently to obtain multiple motion trajectories, thereby determining the reaction time corresponding to each simulation. The average value of the reaction time corresponding to each simulation was determined as the value of the force constant k. i The corresponding reaction time.

[0059] In some embodiments, a relationship between the reaction rate constant information after applying a bias potential energy and the reaction rate constant information can be established based on transition state theory:

[0060]

[0061] in, Information on the reaction rate constant after applying a bias potential. For reaction rate constant information, k B Where is Boltzmann's constant, and T is the thermodynamic temperature. This represents the energy barrier of the first bias potential energy surface.

[0062]

[0063] in, x is the position of the reaction coordinates in the transition state, and x0 is the position of the reaction coordinates in the reactant state. The aforementioned... Substituting the formula into the relationship between the reaction rate constant information after applying the bias potential energy and the actual reaction rate constant information, and taking the logarithm, we fit a straight line with the force constant as the independent variable and the corresponding logarithm of the reaction rate constant information after applying the bias potential energy as the dependent variable. The intercept of this straight line is the logarithm of the reaction rate constant information. Here, the reaction rate constant information after applying the bias potential energy is the reciprocal of the previously calculated reaction time. For example, taking the base-10 logarithm, we have:

[0064]

[0065] With the force constant k as the independent variable, Let be the dependent variable, and fit the line y = αk + β, where

[0066] Among them, T off The reaction time is defined above. The corresponding intercept β is determined through linear fitting, thereby determining the reaction rate constant information. This reaction rate constant information is used to characterize the intrinsic speed of the chemical reaction. Those skilled in the art should understand that the logarithm base 10 mentioned above is merely an example; any other logarithm can be applied to this application, and this application does not limit it.

[0067] In some embodiments, the reaction-related information includes information on the change in reaction free energy; step S15 includes: device 1 applying a second bias potential energy matching the values ​​of each coupling parameter to the reaction coordinates corresponding to the biological system based on different coupling parameter values, performing molecular dynamics simulation using the machine learning force field, and calculating the partial derivative of the Hamiltonian of the biological system with respect to the coupling parameters under each coupling parameter value; and determining the information on the change in reaction free energy using a numerical integration method based on the partial derivative of the Hamiltonian of the biological system with respect to the coupling parameters under each coupling parameter value.

[0068] In some embodiments, the applied second bias potential energy U(x) = k big ·(x-(x0+(x off -x0)·λ)) 2 Where x is the reaction coordinate, x off The position of the reaction coordinates after the reaction occurs (old chemical bonds break and new chemical bonds form) (i.e., the product state) is a constant, while x0 is the position of the reaction coordinates in the reactant state and is also a constant. bigA sufficiently large spring constant is used to constrain the reaction progress within (x0 + (x...)) during molecular dynamics simulations. off -x0)·λ). λ is a coupling parameter, λ∈[0,1]. When λ=0, the reaction progress is constrained to the reactant state, and when λ=1, the reaction progress is constrained to the product state.

[0069] Based on the aforementioned formula for the second bias potential energy, for the set coupling parameter values, the corresponding second bias potential energy is applied to the reaction coordinates. Combined with the machine learning force field, the aforementioned molecular dynamics simulation is performed to obtain the trajectories of each atom in the biological system. These trajectories are recorded frame by frame, each frame including the coordinates, forces acting on each atom in the biological system at the corresponding time, and the total energy of the biological system. In some embodiments, this total energy can be considered as the Hamiltonian of the corresponding biological system. For the partial derivative of the Hamiltonian of the biological system relative to the coupling parameter values, the partial derivative of the Hamiltonian relative to the coupling parameter for each frame obtained from the molecular dynamics simulation can be calculated using appropriate analytical differentiation tools or numerical differentiation methods. Furthermore, the average value of the partial derivatives calculated in each frame is determined as the partial derivative of the Hamiltonian of the biological system relative to the coupling parameter for that coupling parameter value. Due to information on the change of reaction free energy When calculating the change in the free energy of this reaction, numerical integration can be used, combined with the partial derivatives of the Hamiltonian of the biological system with respect to the coupling parameters for each value of the coupling parameters obtained above.

[0070] For the reactions involved in the biological system, the machine learning force field trained in this scheme can conveniently calculate the accurate Hessian matrix, which can be used for transition state search and reaction energy barrier prediction. In some embodiments, the method further includes: step S16 (not shown), where device 1 determines initial transition state structure information based on the structural information corresponding to reactants and products; based on the initial transition state structure information, iterative optimization of the transition state structure is performed using the Hessian matrix calculated by the machine learning force field to determine the optimized transition state structure information.

[0071] For example, for the reactants and products involved in the reaction, the initial transition state structure information can be determined based on their structure using geodesic methods or two-ended methods (e.g., Nudged Elastic Band (NEB) algorithm, Quadratic Synchronous Transport (QST) method). Based on this initial transition state structure information, iterative optimization is performed using the Hessian matrix. This iterative optimization process involves calculating the corresponding Hessian matrix using the machine learning force field trained in this scheme, based on the initial transition state structure. The Hessian matrix is ​​the second derivative of the total energy with respect to the atomic coordinates. It can be determined using the total energy information and the force information of each atom provided by the machine learning force field, combined with appropriate numerical / analytical differentiation methods. Alternatively, it can be processed using appropriate automatic differentiation tools (e.g., the grad function in PyTorch or TensorFlow). Based on this Hessian matrix, the corresponding displacement step size is determined using the Rational Function Optimization (RFO) method or the Partitioned Rational Function Optimization (PRF) method. Based on the displacement step size, the transition state structure is updated, and a convergence check is performed (e.g., the gradient norm, displacement norm, and / or the total energy change of the transition state structure before and after the update is less than the corresponding convergence threshold). If convergence fails, the Hessian matrix is ​​recalculated based on the new transition state structure, and the steps of determining the displacement step size and updating the transition state structure are repeated. If convergence is achieved, the updated transition state structure is output as the transition state optimization structure information.

[0072] In some embodiments, the method further includes: step S17 (not shown), performing intrinsic reaction coordinate optimization based on the transition state optimized structural information to determine the verification result information corresponding to the transition state optimized structural information. For example, based on the transition state optimized structural information, the structural information of the reactants and products calculated by IRC can be obtained through intrinsic reaction coordinate (IRC) optimization. Based on the structural information of the reactants and products calculated by IRC, it is compared with the aforementioned structural information corresponding to the actual reactants and products to determine whether the transition state optimized structural information determined in step S16 is accurate. Since molecules may have multiple conformations, before comparison, the aforementioned structural information can be converted into standard strings (e.g., converting the aforementioned structural information into Canonical SMILES or InChI format strings). The standard string corresponding to the structural information of the reactants calculated by IRC is compared with the standard string corresponding to the structural information of the actual reactants, and the standard string corresponding to the structural information of the products calculated by IRC is compared with the standard string corresponding to the structural information of the actual products. If the matching degree of the structures of the reactants and products is greater than the corresponding threshold, the verification is passed, and the transition state optimized structural information is accurate. Conversely, if the verification fails, the transition state optimization structure information is inaccurate.

[0073] In some embodiments, the method further includes: step S18 (not shown), whereby device 1 determines the corresponding reaction energy barrier information. In some embodiments, the transition state optimized structure information and the total energy information corresponding to the reactants can be determined using the machine learning force field obtained from the aforementioned training, and the difference between the two is the reaction energy barrier information. In some embodiments, to determine accurate reaction energy barrier information, after the aforementioned step S17 has passed verification, the reaction energy barrier information can be determined by subtracting the transition state energy information corresponding to the transition state optimized structure information obtained from the aforementioned IRC optimization and the reactant energy information.

[0074] Figure 2This diagram illustrates a device structure for constructing a machine learning force field for simulating a non-equilibrium biological system according to an embodiment of this application. The device 1 includes a first module 11, a second module 12, and a third module 13. The first module 11 constructs a quantum mechanical simulation model corresponding to the biological system, wherein the quantum mechanical simulation model includes a quantum mechanical cluster model or a combined model of quantum mechanics and molecular dynamics. The second module 12, based on the quantum mechanical simulation model, obtains a corresponding non-equilibrium dataset through transition state scanning, wherein the non-equilibrium dataset includes the structural information of the biological system at each reaction coordinate and the corresponding system state information. The third module 13, based on the non-equilibrium dataset and the equilibrium dataset, trains and obtains a corresponding machine learning force field, wherein the equilibrium dataset includes the structural information of the equilibrium conformation corresponding to the biological system and the corresponding system state information. Here, the... Figure 2 The specific implementation methods corresponding to the first module 11, the second module 12 and the third module 13 shown are the same as or similar to the specific embodiments of the aforementioned steps S11, S12 and S13, and therefore will not be repeated here. They are included here by reference.

[0075] In some embodiments, the first-third module 13 includes a first-third-first unit 131 (not shown) and a first-third-second unit 132 (not shown). The first-third-first unit 131 determines a corresponding training dataset based on the non-equilibrium dataset and the equilibrium dataset, wherein the data from the non-equilibrium dataset and the data from the equilibrium dataset in the training dataset satisfy a preset quantity ratio, and the training dataset includes corresponding structural information and corresponding system state information; the first-third-second unit 132 trains and obtains a corresponding machine learning force field based on the training dataset. Here, the specific implementations of the first-third-first unit 131 and the first-third-second unit 132 are the same as or similar to the specific embodiments of steps S131 and S132 described above, and therefore will not be repeated here, but are incorporated herein by reference.

[0076] In some embodiments, the three-module 13 further includes a three-three unit 133 (not shown). The three-three unit 133, based on the non-equilibrium dataset and the equilibrium dataset, determines other corresponding training datasets in other preset proportions; based on the other training datasets, it trains and obtains corresponding machine learning force fields; based on the machine learning force fields trained from the training datasets obtained in each preset proportion, it selects the final machine learning force field. Here, the specific implementation of the three-three unit 133 is the same as or similar to the specific embodiment of the aforementioned step S133, and therefore will not be repeated, but is incorporated herein by reference.

[0077] In some embodiments, the device 1 further includes a quad module 14 (not shown). The quad module 14 performs molecular dynamics simulations based on the biological system to obtain multiple corresponding frame information; uses a molecular segmentation algorithm to determine multiple structural information corresponding to each frame information, and uses quantum chemical methods to calculate the system state information corresponding to each structural information; based on the structural information and the corresponding system state information, it determines the corresponding equilibrium state dataset. Here, the specific implementation of the quad module 14 is the same as or similar to the specific embodiment of the aforementioned step S14, and therefore will not be repeated, but is included herein by reference.

[0078] In some embodiments, the device 1 further includes a five-module 15 (not shown). The five-module 15, based on an enhanced sampling method, utilizes the machine learning force field to perform molecular dynamics simulations to determine reaction-related information of the chemical reactions corresponding to the biological system. Here, the specific implementation of the five-module 15 is the same as or similar to the specific embodiment of step S15 described above, and therefore will not be repeated here, but is incorporated herein by reference.

[0079] In some embodiments, the device 1 further includes a six-module 16 (not shown). The six-module 16 determines initial transition state structural information based on the structural information corresponding to the reactants and products; based on the initial transition state structural information, iterative optimization of the transition state structure is performed using the Hessian matrix calculated by the machine learning force field to determine the optimized transition state structural information. Here, the specific implementation of the six-module 16 is the same as or similar to the specific embodiment of the aforementioned step S16, and therefore will not be repeated, but is incorporated herein by reference.

[0080] In some embodiments, the device 1 further includes a seven-module 17 (not shown). The seven-module 17 performs intrinsic reaction coordinate optimization based on the transition state optimization structure information to determine the verification result information corresponding to the transition state optimization structure information. Here, the specific implementation of the seven-module 17 is the same as or similar to the specific embodiment of the aforementioned step S17, and therefore will not be repeated here, but is incorporated herein by reference.

[0081] In some embodiments, the device 1 further includes an 8-module 18 (not shown). The 8-module 18 determines the corresponding reaction barrier information. Here, the specific implementation of the 8-module 18 is the same as or similar to the specific embodiment of the aforementioned step S18, and therefore will not be repeated, but is incorporated herein by reference.

[0082] Figure 3 Exemplary systems that can be used to implement the various embodiments described in this application are shown; such as Figure 3As shown in some embodiments, system 300 can function as any of the devices described in each of the embodiments. In some embodiments, system 300 may include one or more computer-readable media having instructions (e.g., system memory or NVM / storage device 320) and one or more processors (e.g., one or more processors 305) coupled to the one or more computer-readable media and configured to execute the instructions to implement the module and thus perform the actions described in this application.

[0083] In one embodiment, the system control module 310 may include any suitable interface controller to provide any suitable interface to at least one of the processors 305 and / or any suitable device or component communicating with the system control module 310.

[0084] The system control module 310 may include a memory controller module 330 to provide an interface to the system memory 315. The memory controller module 330 may be a hardware module, a software module, and / or a firmware module.

[0085] System memory 315 can be used, for example, to load and store data and / or instructions for system 300. In one embodiment, system memory 315 may include any suitable volatile memory, such as suitable DRAM. In some embodiments, system memory 315 may include double data rate type quad synchronous dynamic random access memory (DDR4 SDRAM).

[0086] In one embodiment, the system control module 310 may include one or more input / output (I / O) controllers to provide interfaces to the NVM / storage device 320 and (one or more) communication interfaces 325.

[0087] For example, NVM / storage device 320 may be used to store data and / or instructions. NVM / storage device 320 may include any suitable non-volatile memory (e.g., flash memory) and / or may include any suitable (one or more) non-volatile storage devices (e.g., one or more hard disk drives (HDDs), one or more optical disc drives (CDs), and / or one or more digital universal optical disc (DVD) drives).

[0088] NVM / storage device 320 may include storage resources that are physically part of a device on which system 300 is mounted, or that can be accessed by the device without necessarily being part of it. For example, NVM / storage device 320 may be accessed via a network through one or more communication interfaces 325.

[0089] One or more communication interfaces 325 may provide the system 300 with an interface to communicate over one or more networks and / or with any other suitable device. The system 300 may wirelessly communicate with one or more components of a wireless network in accordance with any of one or more wireless network standards and / or protocols.

[0090] In one embodiment, at least one of the processors 305 may be logically packaged with one or more controllers of the system control module 310 (e.g., memory controller module 330). In one embodiment, at least one of the processors 305 may be logically packaged with one or more controllers of the system control module 310 to form a system-in-package (SiP). In one embodiment, at least one of the processors 305 may be integrated with the logic of one or more controllers of the system control module 310 on the same die. In one embodiment, at least one of the processors 305 may be integrated with the logic of one or more controllers of the system control module 310 on the same die to form a system-on-a-chip (SoC).

[0091] In various embodiments, system 300 may be, but is not limited to, a server, workstation, desktop computing device, or mobile computing device (e.g., laptop computing device, handheld computing device, tablet computer, netbook, etc.). In various embodiments, system 300 may have more or fewer components and / or different architectures. For example, in some embodiments, system 300 includes one or more cameras, a keyboard, a liquid crystal display (LCD) screen (including a touchscreen display), a non-volatile memory port, multiple antennas, a graphics chip, an application-specific integrated circuit (ASIC), and a speaker.

[0092] In addition to the methods and devices described in the above embodiments, this application also provides a computer-readable storage medium storing computer code that, when executed, performs the method described in any of the preceding embodiments.

[0093] This application also provides a computer program product that, when executed by a computer device, performs the method described in any of the preceding claims.

[0094] This application also provides a computer device, the computer device comprising:

[0095] One or more processors;

[0096] Memory, used to store one or more computer programs;

[0097] When the one or more computer programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method as described in any of the preceding methods.

[0098] It should be noted that this application can be implemented in software and / or a combination of software and hardware, for example, using an application-specific integrated circuit (ASIC), a general-purpose computer, or any other similar hardware device. In one embodiment, the software program of this application can be executed by a processor to implement the steps or functions described above. Similarly, the software program of this application (including related data structures) can be stored in a computer-readable recording medium, such as RAM memory, magnetic or optical drives, floppy disks, and similar devices. Furthermore, some steps or functions of this application can be implemented in hardware, for example, as circuitry that cooperates with a processor to perform the various steps or functions.

[0099] Furthermore, a portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0100] Communication media include media through which communication signals containing, for example, computer-readable instructions, data structures, program modules, or other data are transmitted from one system to another. Communication media can include guided transmission media (such as cables and wires (e.g., optical fibers, coaxial cables, etc.)) and wireless (unguided transmission) media capable of propagating energy waves, such as sound, electromagnetic, RF, microwave, and infrared. Computer-readable instructions, data structures, program modules, or other data can be embodied as modulated data signals in, for example, wireless media (such as carrier waves or similar mechanisms embodied as part of spread spectrum technology). The term "modulated data signal" refers to a signal whose one or more characteristics are altered or set in a manner that encodes information in the signal. Modulation can be analog, digital, or a hybrid modulation technique.

[0101] By way of example and not limitation, computer-readable storage media may include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. For example, computer-readable storage media include, but are not limited to, volatile memories such as random access memory (RAM, DRAM, SRAM); and non-volatile memories such as flash memory, various read-only memories (ROM, PROM, EPROM, EEPROM), magnetic and ferromagnetic / ferroelectric memories (MRAM, FeRAM); and magnetic and optical storage devices (hard disks, magnetic tapes, CDs, DVDs); or other media now known or hereafter developed capable of storing computer-readable information / data for use by a computer system.

[0102] Herein, one embodiment of this application includes an apparatus comprising a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the apparatus is triggered to run a method and / or technical solution based on the foregoing embodiments of this application.

[0103] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within this application. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in the apparatus claims may also be implemented by a single unit or device in software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any particular order.

Claims

1. A method for constructing machine learning force fields for simulating non-equilibrium biological systems, wherein, The method includes: Construct a quantum mechanical simulation model corresponding to a biological system, wherein the quantum mechanical simulation model includes a quantum mechanical cluster model or a combined model of quantum mechanics and molecular dynamics; Based on the quantum mechanical simulation model, a corresponding non-equilibrium state dataset is obtained through transition state scanning. The non-equilibrium state dataset includes the structural information of the biological system and the corresponding system state information under each reaction coordinate. Based on the non-equilibrium dataset and the equilibrium dataset, a corresponding machine learning force field is trained and obtained, wherein the equilibrium dataset includes the structural information of the equilibrium conformation corresponding to the biological system and the corresponding system state information; The method further includes: based on the enhanced sampling method, using the machine learning force field to perform molecular dynamics simulation to determine the reaction-related information of the chemical reaction corresponding to the biological system; The reaction-related information includes reaction rate constant information; the reaction-related information for determining the chemical reaction corresponding to the biological system by performing molecular dynamics simulation using the machine learning force field based on the enhanced sampling method includes: Based on different force constant values, a first bias potential energy matching each force constant value is applied to the reaction coordinates corresponding to the biological system. Molecular dynamics simulation is performed using the machine learning force field to determine the reaction time of the biological system under each force constant value. Based on the transition state theory, with the force constant as the independent variable and the logarithm of the reaction rate constant information after applying the bias potential energy as the dependent variable, a straight line is fitted, and the intercept obtained is the logarithm of the reaction rate constant information, thereby determining the reaction rate constant information.

2. The method according to claim 1, wherein, Based on the quantum mechanical simulation model, a corresponding non-equilibrium state dataset is obtained through transition state scanning. This non-equilibrium state dataset includes the structural information and corresponding system state information of the biological system at each reaction coordinate. Based on the quantum mechanical simulation model, a transition state scan is performed to obtain the corresponding scan data, wherein the scan data includes the atomic coordinate information, atomic force information and total energy information of the biological system corresponding to each reaction coordinate during the transition state scan. The scanned data is preprocessed to obtain the non-equilibrium dataset.

3. The method according to claim 1 or 2, wherein, The process of training and obtaining the corresponding machine learning force field based on the non-equilibrium dataset and the equilibrium dataset includes: Based on the non-equilibrium dataset and the equilibrium dataset, a corresponding training dataset is determined, wherein the data from the non-equilibrium dataset and the data from the equilibrium dataset in the training dataset meet a preset quantity ratio, and the training dataset includes corresponding structural information and corresponding system state information; Based on the training dataset, the corresponding machine learning force field is obtained through training.

4. The method according to claim 3, wherein, The step of training and obtaining the corresponding machine learning force field based on the training dataset includes: Based on the structural information in the training dataset, the corresponding system state prediction information is obtained by using machine learning force fields. Based on the system state prediction information and the system state information corresponding to the structural information in the training dataset, and combined with the composite loss function that matches the system state information, the corresponding composite loss is determined. Based on the composite loss, and combined with the corresponding optimization algorithm, the machine learning force field is updated.

5. The method according to claim 3, wherein, The step of training and obtaining the corresponding machine learning force field based on the non-equilibrium dataset and the equilibrium dataset further includes: Based on the aforementioned unbalanced dataset and balanced dataset, other corresponding training datasets are determined in other preset proportions; Based on the other training datasets, the corresponding machine learning force field is obtained through training. The machine learning force field is obtained by training the training dataset based on the preset proportions, and the final machine learning force field is selected from it.

6. The method according to claim 1, wherein, The method further includes: Based on the biological system, molecular dynamics simulations were performed to obtain multiple corresponding frame information. Using a molecular cutting algorithm, multiple structural information corresponding to each frame of information is determined, and quantum chemical methods are used to calculate the system state information corresponding to each structural information. Based on the structural information and the corresponding system state information, the corresponding equilibrium state dataset is determined.

7. The method according to claim 1, wherein, The reaction-related information includes information on changes in reaction free energy; The method based on enhanced sampling, utilizing the machine learning force field to perform molecular dynamics simulations, and determining reaction-related information for the corresponding chemical reactions in the biological system includes: Based on different coupling parameter values, a second bias potential energy matching each coupling parameter value is applied to the reaction coordinates corresponding to the biological system. Molecular dynamics simulation is performed using the machine learning force field to calculate the partial derivative of the Hamiltonian of the biological system with respect to the coupling parameter under each coupling parameter value. Based on the partial derivatives of the Hamiltonian of the biological system with respect to the coupling parameters under various coupling parameter values, the change information of reaction free energy is determined using numerical integration methods.

8. The method according to claim 1, wherein, The method further includes: Based on the structural information corresponding to the reactants and products, the initial transition state structural information is determined; Based on the initial transition state structure information, the Hessian matrix calculated by the machine learning force field is used to perform iterative optimization of the transition state structure, thereby determining the optimized transition state structure information.

9. The method according to claim 8, wherein, The method further includes: Based on the transition state optimization structure information, intrinsic reaction coordinate optimization is performed to determine the verification result information corresponding to the transition state optimization structure information.

10. The method according to claim 8 or 9, wherein, The method further includes: Determine the relevant reaction energy barrier information.

11. A computer device for constructing machine learning force fields for simulating non-equilibrium biological systems, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 10.

12. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 10.

13. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method as described in any one of claims 1 to 10.

Images