Adversarial training of deep force field models for stable simulation

Through the adversarial training method, the data set is expanded using adversarial conformation, and the instability problem of the deep force field model under extreme stress conditions is solved, achieving the continuous stability of molecular dynamics simulation.

CN118866181BActive Publication Date: 2024-11-29XI AN JIAOTONG UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411330354.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-24
Publication Date
2024-11-29
Estimated Expiration
2044-09-24

AI Technical Summary

Technical Problem

The existing deep force field model is easy to predict extreme stress conditions in molecular dynamics simulation, resulting in unstable simulation results and the inability to achieve sustainable and stable simulations.

Method used

By creating molecular data sets, augmenting the data sets using adversarial sample generation algorithms, introducing adversarial conformations, and performing adversarial training to enhance the robustness of the model and ensuring the stability of the prediction results in extreme cases.

Benefits of technology

It effectively avoids the collapse of the simulation process, improves the stability and reliability of the model when predicting extreme forces, and ensures the continuous stability of molecular dynamics simulation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118866181B_ABST
    Figure CN118866181B_ABST
Patent Text Reader

Abstract

The present application discloses an adversarial training method for a deep force field model for stable simulation, which belongs to the field of molecular dynamics, and includes: establishing a molecular data set; training the initial model by using the molecular data set as the first training data and the first loss function to obtain a pre-trained model; inputting the molecular data set into the pre-trained model to obtain the predicted dynamic state of the molecules in the molecular data set, and expanding the molecular data set into an adversarial conformational data set using an adversarial sample generation algorithm according to the predicted dynamic state; training the pre-trained model by using the adversarial conformational data set as the second training data and the second loss function to obtain a deep force field model. This solves the problem that the deep mechanics model based on machine learning cannot achieve continuous and stable molecular dynamics simulation due to the prediction of extreme force conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application belongs to the field of molecular dynamics, and specifically relates to an adversarial training method for a deep force field model for stable simulation, and a molecular dynamics simulation method, apparatus, storage medium, device and computer program product. Background Art

[0002] In the interdisciplinary research between artificial intelligence and chemistry, deep learning techniques are increasingly being used in molecular dynamics simulations. These simulations are essential for understanding and predicting molecular behavior, especially in areas such as drug design, materials science, and energy development. By simulating the behavior of molecules under different conditions, researchers can reveal the complex relationship between molecular structure and its function, thereby promoting the discovery of new materials and drugs.

[0003] Currently, in order to realize molecular dynamics simulation, researchers have developed a force field model based on machine learning, which predicts the properties and dynamic state of molecules by learning the relationship between molecular structure and potential energy.

[0004] However, when the simulation trajectory enters a non-physical state, the model may predict extreme force conditions, resulting in unstable simulation results, which in turn limits the robustness of the model in predicting unknown molecular conformations. Therefore, existing deep mechanics models often cannot achieve continuous and stable simulations, ultimately causing the simulation process to collapse. Summary of the invention

[0005] The present application aims to provide an adversarial training method for a deep force field model for stable simulation, as well as a molecular dynamics simulation method, apparatus, storage medium, equipment and computer program product, which at least solves the problem that the deep mechanics model cannot achieve continuous and stable simulation due to the prediction of extreme force conditions during molecular dynamics simulation using a force field model based on machine learning.

[0006] In a first aspect, an embodiment of the present application discloses an adversarial training method for a deep force field model for stable simulation, comprising:

[0007] Establishing a molecular data set; the molecular data set comprises an original conformation of each of the plurality of molecules; the original conformation is used to characterize the actual dynamic state of the molecule;

[0008] Using the molecular data set as first training data, and training the initial model using the first training data and a first loss function to obtain a pre-trained model;

[0009] Inputting the molecular data set into the pre-trained model to obtain the predicted dynamic state of the molecules in the molecular data set, and expanding the molecular data set into an adversarial conformation data set using an adversarial sample generation algorithm according to the predicted dynamic state; the adversarial conformation data set includes the original conformation of the molecule and the adversarial conformation; the adversarial conformation is used to characterize the dynamic state of the molecule under mechanical perturbation;

[0010] Using the adversarial conformation data set as second training data, and training the pre-trained model using the second training data and a second loss function to obtain a deep force field model;

[0011] The adversarial sample generation algorithm determines the adversarial conformation of the molecule through the following formula:

[0012] ;

[0013] in, Represents the updated spatial coordinate value of the molecule; r Represents the actual spatial coordinate value of the molecule in the molecular data set; represents the identity matrix; Represents the preset disturbance coefficient, which is a scalar used to represent the magnitude of the disturbance; represents a symbolic function; Indicates the gradient of the spatial coordinate value; J 1 represents the first loss function; θ represents the preset weight parameter; X represents a molecule; y Represents molecules X The actual dynamic state.

[0014] In a second aspect, the present application also discloses a molecular dynamics simulation method, including:

[0015] Acquire an experimental data set to be simulated; the experimental data set includes a plurality of original conformations of the molecule; the original conformations are used to characterize the actual dynamic state of the molecule;

[0016] The experimental data set is input into a deep force field model trained by the adversarial training method of the deep force field model for stable simulation as described in the first aspect to obtain the predicted dynamic states of the molecules in the experimental data set as the simulation results of molecular dynamics.

[0017] In a third aspect, the embodiment of the present application further discloses an adversarial training device for a deep force field model for stable simulation, comprising:

[0018] An initial training set module is used to establish a molecular data set; the molecular data set includes an original conformation of each of the multiple molecules; the original conformation is used to characterize the actual dynamic state of the molecule;

[0019] A pre-training module, used to use the molecular data set as first training data, and train the initial model through the first training data and a first loss function to obtain a pre-training model;

[0020] An adversarial conformation set module, used for inputting the molecular data set into the pre-trained model to obtain the predicted dynamic state of the molecules in the molecular data set, and expanding the molecular data set into an adversarial conformation data set according to the predicted dynamic state by using an adversarial sample generation algorithm; the adversarial conformation data set includes the original conformation of the molecule and the adversarial conformation; the adversarial conformation is used to characterize the dynamic state of the molecule under mechanical perturbation;

[0021] The final training module is used to use the adversarial conformation data set as second training data, and train the pre-training model through the second training data and a second loss function to obtain a deep force field model.

[0022] In a fourth aspect, the present application also discloses a molecular dynamics simulation device, comprising:

[0023] A data preparation module, used to obtain an experimental data set to be simulated; the experimental data set includes a plurality of original conformations of the molecule; the original conformations are used to characterize the actual dynamic state of the molecule;

[0024] A data simulation module is used to input the experimental data set into a deep force field model trained by the adversarial training method of the deep force field model for stable simulation as described in the first aspect, so as to obtain the predicted dynamic state of the molecules in the experimental data set as the simulation result of molecular dynamics.

[0025] In a fifth aspect, an embodiment of the present application further discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps described in the first aspect or the second aspect are implemented.

[0026] In a sixth aspect, an embodiment of the present application further discloses an electronic device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps described in the first aspect or the second aspect.

[0027] In a seventh aspect, an embodiment of the present application further discloses a computer program product, on which a computer program is stored, and when the computer program is executed by a processor, the steps described in the first aspect or the second aspect are implemented.

[0028] In summary, in the embodiments of the present application, the deep force field model is trained by the first loss function to ensure the basic ability of the model to predict molecular properties and dynamic states, and then the molecular data set is input into the pre-trained model and the prediction of the dynamic state is obtained through the adversarial sample generation algorithm, so as to further introduce adversarial conformations and non-extremeize the stress conditions in extreme cases, thereby ensuring the stability and reliability of the prediction results. Finally, the data set containing the adversarial conformation is trained by the second loss function, so that the deep force field model finally obtained can remain stable when predicting extreme forces, thereby avoiding the collapse of the simulation process. Therefore, the method based on the embodiments of the present application solves the problem in the related art that when performing molecular dynamics simulations through a force field model based on machine learning, extreme stress conditions will be predicted, thereby causing the deep mechanics model to be unable to achieve continuous and stable molecular dynamics simulations. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0030] Figure 1 This embodiment provides an adversarial training method for a deep force field model for stable simulation;

[0031] Figure 2 is a schematic diagram of the antagonistic conformation in the examples of the present application;

[0032] Figure 3 This is another adversarial training method for a deep force field model for stable simulation provided by this embodiment;

[0033] Figure 4 It is a molecular dynamics simulation method provided in this embodiment;

[0034] Figure 5 This is an experimental process of adversarial training of a deep force field model for stable simulation provided by an embodiment of the present application;

[0035] Figure 6 is a schematic diagram of the structure of an adversarial training device for a deep force field model for stable simulation provided in this embodiment;

[0036] Figure 7is a schematic structural diagram of a molecular dynamics simulation device provided in this embodiment;

[0037] Figure 8 is a block diagram of an electronic device provided in an embodiment of the present application;

[0038] Fig. 9 It is a block diagram of another electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] The exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0040] Figure 1 An adversarial training method for a deep force field model for stable simulation provided by an embodiment of the present application may include the following steps:

[0041] Step 101, establishing a molecular data set.

[0042] The molecular data set includes the original conformation of each molecule in a plurality of molecules; the original conformation is used to characterize the actual dynamic state of the molecule.

[0043] In some embodiments of the present application, it is first necessary to establish a molecular data set, which contains the original conformations of multiple molecules. These original conformations are the training data of the molecular dynamics simulation model, which are used to characterize the actual dynamic state of the molecule in the experiment or theoretical calculation. In this step, the construction of the molecular data set is to provide enough samples to ensure that the deep force field model can learn the behavior and properties of the molecule in different states.

[0044] For example, in order to establish a molecular dataset for training a deep force field model, a series of molecular structural information can be collected from public chemical databases, including the atomic coordinates, electronic structure, and related dynamic parameters (such as force conditions and energy) of the molecules. These data are usually obtained through high-precision quantum chemical calculation methods, such as density functional theory (DFT). These molecular data can then be divided into training sets, validation sets, and test sets in a certain proportion for use in the model training process. In order to ensure the uniqueness of the data, the uniqueness of the molecular types in the molecular dataset can be maintained as much as possible during training. In this way, the molecular dataset can provide the model with rich learning resources to predict the behavior of molecules under different environmental conditions.

[0045] Step 102: Use the molecular data set as first training data, and train the initial model using the first training data and a first loss function to obtain a pre-trained model.

[0046] In some embodiments of the present application, a molecular dataset is used as training data to train a deep force field model through a defined first loss function. The purpose of this training process is to obtain a pre-trained model that can make basic predictions about the dynamic state of molecules. The first loss function is usually designed to evaluate the difference between the molecular properties (such as energy and force) predicted by the model and the actual values, thereby guiding the optimization of model parameters.

[0047] For example, you can first select a dataset containing multiple conformations of water molecules and then define the first loss function. By minimizing this loss function, the model learns how to accurately predict the energy and forces of the molecule, thereby obtaining a pre-trained model. In this example, the dataset may contain extreme conformations of the molecule and their corresponding DFT calculation results. In this way, the pre-trained model can capture the behavioral characteristics of the molecule in different environments.

[0048] Step 103, inputting the molecular data set into the pre-training model to obtain the predicted kinetic states of the molecules in the molecular data set, and expanding the molecular data set into an adversarial conformational data set based on the predicted kinetic states using an adversarial sample generation algorithm.

[0049] Among them, the adversarial conformation data set includes the original conformation and adversarial conformation of the molecule; the adversarial conformation is used to characterize the dynamic state of the molecule under mechanical perturbation.

[0050] In some embodiments of the present application, it is necessary to input the molecular data set into the pre-trained model to obtain a prediction of the dynamic state of the molecules in the data set. The adversarial conformation is the conformation after the original molecular conformation is subjected to an adversarial disturbance. The adversarial disturbance has the characteristics of small disturbance amount and can cause the model to produce significant prediction deviations. The core of this step is to use the pre-trained model to analyze the original conformation of the molecule and predict its possible reaction when subjected to mechanical disturbance. Based on these predictions, an adversarial sample generation algorithm can be used to create an extended data set containing adversarial conformations, which represent the molecular mechanical states that the current model cannot correctly analyze.

[0051] like Figure 2As shown, in the embodiments of the present application, the adversarial conformation refers to a new conformation (B, C, D) generated by introducing a small perturbation in a certain conformation of the molecule (A in the figure). These perturbations are usually based on the gradient information predicted by the model and are of limited size. The purpose is to estimate the impact of molecular structural changes on model predictions. The adversarial conformation does not necessarily correspond to the conformation that the molecule will spontaneously form under natural conditions, but is used to test and improve the robustness of the model in predicting molecular behavior.

[0052] For example, in the process of studying a dynamics dataset of a water molecule under certain physical conditions. By inputting the molecular conformation in the dynamics data into the pre-trained model, adversarial conformations are generated using adversarial sample generation methods. These conformations may include the conformations of water molecules in extreme dehydration or compression states, and the DFT method is used to calculate the forces on the adversarial conformations. These adversarial conformations and corresponding forces are then added to the original dynamics dataset to form an extended adversarial conformation dataset. This extended dataset not only contains the normal conformation of the molecule (i.e., the data in the molecular dataset), but also adversarial conformations that may appear under perturbations. These adversarial conformations represent extreme molecular mechanical states that the current model cannot correctly analyze.

[0053] In the embodiments of the present application, the adversarial sample generation algorithm specifically refers to the process of obtaining the molecular conformation through the following formula:

[0054] ;

[0055] in, Represents the updated spatial coordinate value of the molecule; r Represents the actual spatial coordinate value of the molecule in the molecular data set; represents the identity matrix; Represents the preset disturbance coefficient, which is a scalar used to represent the magnitude of the disturbance; represents a symbolic function; Indicates the gradient of the spatial coordinate value; J 1 represents the first loss function; θ represents the preset weight parameter; X represents a molecule; y Represents molecules X The actual dynamic state.

[0056] The advantage of using this process is that it simulates the influence of external conditions on the spatial coordinates of molecules by introducing a small disturbance. This method can help the model learn how to predict the behavior of molecules under different environmental conditions. The introduction of disturbance is based on the direction calculation of the gradient of the loss function, which increases the loss function. The adversarial sample obtained in this way has a destructive effect, causing the calculation results to deviate from the prediction range of the current model. In addition, it should be emphasized that in the process of the above embodiment, the use of a sign function ensures that the size of the disturbance is uniform, which helps to avoid the expected deviation that may be caused by excessive disturbances.

[0057] Step 104 , using the adversarial conformation data set as second training data, and training the pre-trained model using the second training data and the second loss function to obtain a deep force field model.

[0058] In some embodiments of the present application, the adversarial conformation dataset is used as training data, and the pre-trained deep force field model is further trained by the second loss function. The purpose of this step is to enhance the robustness of the model through adversarial training, so that it can be more stable when predicting molecular dynamic states. The second loss function is specially designed to evaluate the predictive performance of the model for adversarial conformations, thereby ensuring that the model can handle extreme or unconventional dynamic states.

[0059] For example, you can first select a dataset containing multiple conformations of water molecules, and then define a second loss function. By minimizing this loss function, the model learns how to accurately predict the energy and forces of the molecule, thereby obtaining a trained deep force field model. In this example, the dataset may contain molecular conformations of water molecules in extremely dehydrated or compressed states.

[0060] In summary, in the embodiments of the present application, the deep force field model is trained by the first loss function to ensure the basic ability of the model to predict molecular properties and dynamic states, and then the molecular data set is input into the pre-trained model and the prediction of the dynamic state is obtained through the adversarial sample generation algorithm, so as to further introduce adversarial conformations and non-extremeize the stress conditions in extreme cases, thereby ensuring the stability and reliability of the prediction results. Finally, the data set containing the adversarial conformation is trained by the second loss function, so that the deep force field model finally obtained can remain stable when predicting extreme forces, thereby avoiding the collapse of the simulation process. Therefore, the method based on the embodiments of the present application solves the problem in the related art that when performing molecular dynamics simulations through a force field model based on machine learning, extreme stress conditions will be predicted, thereby causing the deep mechanics model to be unable to achieve continuous and stable molecular dynamics simulations.

[0061] Figure 3This is another adversarial training method for a deep force field model for stable simulation provided by this embodiment. The method may include the following steps:

[0062] Step 201, establishing a molecular data set.

[0063] The molecular data set includes the original conformation of each molecule in a plurality of molecules; the original conformation is used to characterize the actual dynamic state of the molecule.

[0064] The method shown in this step has been explained in step 101 and will not be repeated here.

[0065] Step 202: Use the molecular data set as first training data, and train the initial model using the first training data and a first loss function to obtain a pre-trained model.

[0066] The method shown in this step has been explained in step 102 and will not be repeated here.

[0067] Optionally, step 202 includes the following sub-steps:

[0068] Sub-step 2021 , inputting the molecular data set into the deep force field model to be trained to obtain the predicted dynamic states of the molecules in the molecular data set.

[0069] In some embodiments of the present application, as a preliminary stage in the deep force field model training process, the model to be trained is used to predict the dynamic state of molecules in the molecular data set. In this stage, the model predicts the behavior and properties of molecules under different conditions by analyzing the information in the molecular data set, such as atomic coordinates and molecular structure.

[0070] For example, suppose researchers are developing a new deep force field model to predict the transport path of a particular drug molecule in the human body. First, a data set containing the original conformation of the drug molecule under different physiological conditions can be collected. This data set may include three-dimensional structural information of the molecule and related kinetic parameters, which are obtained through high-precision quantum chemical calculation methods. Then, this data can be input into the deep force field model to be trained. The model analyzes this data and predicts various kinetic states that the drug molecule may encounter in the body. This step provides a preliminary training basis for the model, enabling it to more accurately predict the behavior of the molecule in subsequent training. In this way, the model is learning how to deal with molecular dynamics problems, laying the foundation for subsequent adversarial training and model optimization.

[0071] Sub-step 2022, establishing a first loss function according to the predicted kinetic states and the actual kinetic states of the molecules in the molecular data set.

[0072] In some embodiments of the present application, a first loss function is established based on the predicted kinetic state and the actual kinetic state of the molecules in the molecular data set. This loss function is a key component of the training process, which measures the difference between the model prediction results and the actual results and guides the optimization of the model parameters. The first loss function usually includes the prediction errors of energy and force, ensuring that the model can accurately predict the behavior of molecules in different environments.

[0073] For example, you can use a dataset containing a variety of small organic molecules whose dynamical states are calculated using DFT. This data can then be fed into a deep force field model, which outputs the predicted energies and forces for the molecules. A first loss function is then established based on the predicted energies and forces. By minimizing this loss function, the model learns how to more accurately predict the energy and forces of the molecules, thereby improving its performance in molecular dynamics simulations.

[0074] Optionally, the dynamic state includes a force value and an energy value, and step 2022 includes the following sub-steps:

[0075] Sub-step 20221, establishing the expected force values ​​of the molecules in the molecular data set according to the predicted force values ​​and the actual force values ​​of the molecules in the molecular data set, and establishing the expected energy values ​​of the molecules in the molecular data set according to the predicted energy values ​​and the actual energy values ​​of the molecules in the molecular data set.

[0076] Among them, the expected value of molecular force is the expectation of the square of the second-order norm of the difference between the predicted force value of the molecule in the molecular data set and the actual force value, and the expected value of molecular energy is the expectation of the square of the second-order norm of the difference between the predicted energy value of the molecule in the molecular data set and the actual energy value.

[0077] In some embodiments of the present application, it is necessary to establish a molecular force expectation value and a molecular energy expectation value, which are based on the difference between the predicted force value and the actual force value of the molecule in the molecular data set, and the predicted energy value and the actual energy value. The molecular force expectation value and the molecular energy expectation value are important indicators for evaluating the accuracy of model prediction, and they are usually obtained by calculating the square of the second-order norm of the difference between the predicted value and the actual value.

[0078] For example, in the process of developing a model to predict the behavior of a new synthetic material molecule. Experimental data of the molecule under various physical conditions, including force values ​​and energy values, are collected. The model predicts the force values ​​and energy values ​​of these molecules under the same conditions. Then the expected value of the molecular force and the expected value of the molecular energy can be calculated. In this way, the accuracy of the model in predicting the molecular behavior of the new synthetic material can be evaluated, and the model parameters can be optimized accordingly. This embodiment shows how to construct a loss function by calculating the expected value to guide the adversarial training process of the deep force field model for stable simulation.

[0079] Sub-step 20222, taking the sum of the expected value of the molecular energy and the expected value of the molecular force as the first loss function.

[0080] In some embodiments of the present application, it is necessary to combine the expected value of molecular energy with the expected value of molecular force to form a first loss function for model training. This loss function is the key in the model training process. It measures the deviation between the energy and force predicted by the model and the experimental data, and guides the optimization of model parameters.

[0081] For example, when developing a model to predict the stability of a drug molecule, the force and energy values ​​of the drug molecule are obtained from the experiment, and these values ​​are predicted by the model. In order to train the model, a first loss function can be established, which is the sum of the expected value of the molecule's energy and the expected value of the molecule's force. By minimizing this loss function, the model can learn how to more accurately predict the energy and force of the molecule in different environments, thereby improving its application performance in drug design and chemical reaction simulation.

[0082] In some embodiments of the present application, according to sub-step 20221 and sub-step 20222, the first loss function obtained may be:

[0083] ;

[0084] in, represents the first loss function, E e represents the expected value of molecular energy, E f represents the expected value of the molecular force, represents the force value predicted by the molecule in the molecular dataset, E({r}) Indicates the actual force value of the molecule in the molecular data set, represents the energy value of the molecule prediction in the molecule dataset, F Represents the actual energy value of the molecule in the molecular data set, To find the sign of the second-order norm.

[0085] in, You can askE({r}) Space coordinates r The partial derivative of is obtained: .

[0086] The advantage of the above loss function is that it balances the prediction errors in both energy and force, allowing the model to fully learn the dynamic properties of the molecule. The combination of energy expectation and force expectation helps the model achieve higher accuracy in predicting the stability and reactivity of the molecule. In addition, by performing a square root operation (second-order norm) on the predictions of energy and force, the loss function gives greater weight to large prediction errors, thereby prompting the model to focus more on reducing these large errors.

[0087] Sub-step 2023, training the deep force field model to be trained by using the first loss function, and obtaining a pre-trained model when the output value of the deep force field model converges.

[0088] In some embodiments of the present application, the deep force field model to be trained is trained using the first loss function until the output value of the model converges, thereby obtaining a pre-trained model. This step ensures that the model can accurately predict the dynamic state of the molecule based on the information in the molecular data set. Convergence is an important indicator in the model training process, which indicates that the model is stable enough to effectively predict new data.

[0089] For example, the difference between the energy and force predicted by the model and the true value calculated by DFT can be calculated, and the model parameters can be iteratively updated through stochastic gradient descent (SGD) or adaptive moment estimation (Adam). During the training process, the value of the loss function will gradually decrease until it reaches a preset threshold or no longer changes significantly, indicating that the model has converged.

[0090] Step 203 , inputting the molecular data set into the pre-trained model to obtain the predicted kinetic states of the molecules in the molecular data set.

[0091] In some embodiments of the present application, it is necessary to predict the dynamic state of molecules in the molecular data set through a deep force field model. In this step, the model uses the structural information of the molecule, such as atomic coordinates and molecular conformation, to predict the behavior of the molecule under different environments and conditions. This process is the basis for the creation of the adversarial conformational data set because it provides the model with a preliminary understanding of the molecular dynamic state.

[0092] For example, researchers may use a dataset containing a variety of protein molecules. The molecular conformations in these datasets are obtained through high-precision quantum chemical calculation methods. Inputting these data into the deep force field model, the model will predict the dynamic state of the protein molecule, including its possible conformational changes and interactions. These predictions will provide the necessary information for determining the antagonistic conformations of the molecule, thereby expanding the dataset in the subsequent steps.

[0093] Step 204 , using the predicted kinetic state as an input value of an adversarial sample generation algorithm to calculate and obtain the adversarial conformation of the molecule.

[0094] In some embodiments of the present application, an adversarial sample generation algorithm is used to determine the adversarial conformations of the molecule based on the predicted kinetic state. This step is a key step in creating an adversarial conformation dataset, which generates molecular conformations that may not naturally occur under actual conditions based on the kinetic state predicted by the model. These adversarial conformations are used to simulate molecular behavior, thereby enhancing the robustness of the model by expanding the dataset.

[0095] For example, a deep force field model can be used to first predict the dynamic state of a molecule. Based on these predictions, adversarial conformations can be generated by adding small perturbations to the normal conformation of the molecule. The generated adversarial conformations may include atypical stretching or angle changes of chemical bonds in the molecule. These adversarial conformations are then added to the original dataset to form an extended dataset containing both original and adversarial conformations. In this way, the model can learn how to deal with molecular behaviors that are difficult to observe in experiments during training, providing a more comprehensive data basis for further training and application of the model.

[0096] Optionally, the dynamic state includes a spatial coordinate value and a force value, and step 204 includes the following sub-steps:

[0097] Sub-step 2041, obtaining the first loss function when the pre-training model training is completed.

[0098] In some embodiments of the present application, a first loss function is obtained when the pre-trained model training is completed. This loss function reflects the accuracy of the model's prediction of the molecular dynamics state during the training process and is used to evaluate the model performance. This loss function is obtained in order to use its gradient information to generate adversarial conformations in subsequent steps.

[0099] For example, in the process of using a deep force field model to predict the charge distribution of a new type of nanomaterial, the model has been pre-trained. Then, this loss function can be obtained after the model training is completed, which will be used to guide the generation of adversarial conformations to improve the model's ability to predict molecular behavior.

[0100] Sub-step 2042, determining the increment of the spatial coordinate value according to the gradient of the first loss function with respect to the spatial coordinate value, so as to update the spatial coordinate value of the molecule.

[0101] In some embodiments of the present application, the gradient of the first loss function to the spatial coordinate value is used to determine the increment of the spatial coordinate value, thereby updating the spatial coordinate value of the molecule. This process is a key step in generating adversarial conformations, which guides small changes in the molecular conformation by analyzing the gradient of the loss function and simulates the impact of external perturbations on the molecular structure.

[0102] For example, when studying the conformation of a protein molecule, the spatial coordinate values ​​of the protein molecule in the initial data have been obtained, and the kinetic state after drug binding has been predicted by the deep force field model. In order to generate an adversarial conformation, the gradient of the first loss function to the spatial coordinate value can be calculated, and the increment of the spatial coordinate value can be determined. This increment is based on the deviation between the model prediction and the experimental data. By applying this increment, the spatial coordinate value of the protein molecule can be updated to generate an adversarial conformation that simulates the conformational change caused by drug binding.

[0103] Sub-step 2043, determining the force value of the molecule after the spatial coordinate value is updated according to the updated spatial coordinate value, and determining the dynamic state of the molecule after the spatial coordinate value and the force value are updated as the antagonistic conformation of the molecule.

[0104] In some embodiments of the present application, the force values ​​of the molecules after the updated spatial coordinate values ​​are determined according to the updated spatial coordinate values, and these updated dynamic states are defined as the antagonistic conformations of the molecules. This step is the final stage of generating antagonistic conformations, which ensures that small changes in the molecular conformation can be reflected in the update of the force values, thereby simulating conformational conditions that are not captured in actual applications.

[0105] For example, in the process of studying the reaction kinetics of a drug molecule, the spatial coordinate values ​​of the molecule have been updated according to the gradient of the spatial coordinate values ​​of the first loss function. Next, the force values ​​of these updated conformations can be determined using quantum chemical calculation methods. These force values, together with the updated spatial coordinate values, constitute the antagonistic conformations of the drug molecule. These antagonistic conformations can be used to further train the deep force field model so that it can predict the behavior of drug molecules under extreme acid-base conditions.

[0106] Step 205 , using the adversarial conformation data set as second training data, and training the pre-trained model using the second training data and the second loss function to obtain a deep force field model.

[0107] The method shown in this step has been described in step 104 and will not be repeated here.

[0108] Optionally, step 205 includes the following sub-steps:

[0109] Sub-step 2051, inputting the adversarial conformational dataset into the pre-trained model to obtain the predicted dynamic states of the molecules in the adversarial conformational dataset.

[0110] In some embodiments of the present application, it is necessary to input the adversarial conformational dataset into the pre-trained model to obtain a prediction of the dynamic state of the molecules in the dataset. This process is the preliminary stage of adversarial training, which aims to use the pre-trained model to predict the expanded dataset, thereby enhancing the processing power of the model.

[0111] For example, in the process of studying the behavior of a new catalyst molecule, a series of adversarial conformations have been generated by adding small perturbations to the normal conformation of the catalyst molecule, and these conformations together with the normal conformation constitute an adversarial conformation dataset. This dataset can then be input into a pre-trained deep force field model. The model analyzes this data and predicts the possible dynamic states of the catalyst molecule, such as changes in atomic spacing, breaking or forming chemical bonds. These prediction results will be used for subsequent adversarial training to improve the accuracy and robustness of the model in practical applications.

[0112] Sub-step 2052, establishing a second loss function based on the predicted dynamic state of the molecule in the adversarial conformational dataset and the actual dynamic state.

[0113] In some embodiments of the present application, a second loss function is established based on the predicted kinetic state and the actual kinetic state of the molecule in the adversarial conformational dataset. This loss function is the core of the adversarial training process, which evaluates the accuracy of the model's prediction for the expanded dataset and guides the further optimization of the model parameters.

[0114] For example, suppose that in the process of developing a model to predict the stability of a drug molecule at different body temperatures, a series of antagonistic conformations have been generated. Then, a second loss function can be established based on the antagonistic conformations, and by minimizing this loss function, the model can learn how to accurately predict the forces on the drug molecule, thereby improving its predictive ability in clinical applications.

[0115] Optionally, the dynamic state includes a force value, and step 2052 includes the following sub-steps:

[0116] Sub-step 20521, establishing the expected value of the antagonistic force of the antagonistic conformational data set according to the predicted force values ​​of the molecules in the antagonistic conformational data set and the actual force values, and using the expected value of the antagonistic force as the second loss function.

[0117] The expected value of the antagonistic force is the expectation of the square of the second-order norm of the difference between the predicted force value of the molecule in the antagonistic conformation dataset and the actual force value.

[0118] In some embodiments of the present application, the purpose is to establish an expected value of adversarial force based on the predicted force values ​​of molecules in the adversarial conformation dataset and the actual force values, and use it as the second loss function. This loss function is used to evaluate the accuracy of the model's prediction of molecular force during adversarial training.

[0119] For example, in the process of studying the chemical stability of a compound, a series of antagonistic conformations have been generated, which simulate various spatial changes of the compound. Then the model can predict the force values ​​of these antagonistic conformations and compare them with the actual force values ​​measured experimentally to determine an expected value. By minimizing this expected value, the model can learn how to more accurately predict the force of the compound, thereby improving its application performance in materials science research.

[0120] In some embodiments of the present application, according to sub-step 20521, the second loss function obtained may be:

[0121] ;

[0122] in, represents the second loss function, represents the expected value of the resisting force, represents the force value predicted by the molecule in the confrontation conformation dataset, F’ represents the actual force value of the molecule in the confrontation conformation dataset, To find the sign of the second-order norm.

[0123] in, Through Find the space coordinates r The partial derivative of is obtained: .

[0124] The second loss function is constructed to specifically evaluate and optimize the accuracy of the model's prediction of molecular forces in adversarial training. The advantage of this loss function is that it directly relates the difference between the force predicted by the model and the actual force, so that the focus of model training is on reducing this difference. This method is particularly suitable for adversarial training.

[0125] Sub-step 2053, training the pre-trained model through the second loss function, and obtaining the trained deep force field model when the output value of the pre-trained model converges.

[0126] In some embodiments of the present application, the pre-trained model is trained using the second loss function until the model output value converges, thereby obtaining a trained deep force field model. This step is a key step to ensure that the model can avoid extreme situations when predicting molecular dynamics states, thereby causing the model calculation to crash.

[0127] For example, suppose a dataset of adversarial conformations of drug molecules is used to train the model. A second loss function is defined. Then, by minimizing this loss function, the model can learn how to accurately predict the forces on drug molecules. During the training process, the value of the loss function will gradually decrease until it reaches a preset threshold or no longer changes significantly, indicating that the model has converged.

[0128] In summary, in the embodiments of the present application, the deep force field model is trained by the first loss function to ensure the basic ability of the model to predict molecular properties and dynamic states, and then the molecular data set is input into the pre-trained model and the prediction of the dynamic state is obtained through the adversarial sample generation algorithm, so as to further introduce adversarial conformations and non-extremeize the stress conditions in extreme cases, thereby ensuring the stability and reliability of the prediction results. Finally, the data set containing the adversarial conformation is trained by the second loss function, so that the deep force field model finally obtained can remain stable when predicting extreme forces, thereby avoiding the collapse of the simulation process. Therefore, the method based on the embodiments of the present application solves the problem in the related art that when performing molecular dynamics simulations through a force field model based on machine learning, extreme stress conditions will be predicted, thereby causing the deep mechanics model to be unable to achieve continuous and stable molecular dynamics simulations.

[0129] Figure 4 This embodiment provides a molecular dynamics simulation method, which includes:

[0130] Step 301: Obtain an experimental data set to be simulated.

[0131] Among them, the experimental data set contains the original conformations of multiple molecules; the original conformations are used to characterize the actual dynamic state of the molecules.

[0132] In some embodiments of the present application, an experimental data set to be simulated is obtained, which contains the original conformations of multiple molecules. These original conformations are the basis for simulating the molecular dynamics state, and they reflect the real behavior and properties of molecules under specific conditions.

[0133] For example, to simulate ion transport in lithium batteries, researchers may collect a series of raw conformational data for lithium ions in different electrolyte environments. These data can be obtained through experimental measurements or high-precision computational methods, such as quantum chemical calculations using DFT. The collected data set should include various possible positions and configurations of lithium ions in the electrolyte, as well as the corresponding energy and force data. Such an experimental data set provides a solid starting point for the simulation, ensuring that the simulation results can reflect the actual physical and chemical processes. In this way, researchers can better understand and predict the behavior of lithium ions during battery operation, thereby providing support for battery design and optimization.

[0134] Step 302, input the experimental data set into the deep force field model trained by the adversarial training method of the deep force field model for stable simulation disclosed in the embodiment of the present application, so as to obtain the predicted dynamic state of the molecules in the experimental data set as the simulation result of molecular dynamics.

[0135] In some embodiments of the present application, the experimental data set is input into a trained deep force field model to obtain a prediction of the dynamic state of the molecules in the data set. This step uses the trained model to predict the behavior of the molecules in the experimental data set, thereby generating simulation results. These results can help researchers understand the behavior of molecules under actual conditions and guide experimental design and theoretical analysis.

[0136] For example, suppose that in the process of studying the transport mechanism of a certain drug molecule in the human body, the original conformational data of the drug molecule under different physiological conditions can be collected and input into the deep force field model that has been trained to enhance stability. Based on these data, the model will predict various dynamic states that the drug molecule may encounter in the body, including interactions with biomacromolecules, transport through cell membranes, etc. Through this simulation, the behavior of drug molecules can be predicted, drug design can be optimized, and the therapeutic effect can be improved.

[0137] like Figure 5 As shown, the complete process from data preparation at the initial stage of training to finally obtaining a stable model according to the method provided by the present application is shown, which specifically includes steps S1 to S6:

[0138] Step S1: Use a standard molecular dataset to train the deep force field model to obtain a pre-trained model. This step is the initial stage of model training, which aims to train the model with known molecular data to prepare for subsequent adversarial training.

[0139] Step S2: Based on the pre-trained model, use the adversarial sample generation algorithm to generate adversarial conformations of the molecule. This step is the process of generating adversarial samples, which simulates situations that may lead to prediction errors by introducing small perturbations based on the pre-trained model.

[0140] Step S3: Use quantum mechanics methods to calculate the forces on the molecular adversarial conformations and obtain adversarial molecular samples. In this step, quantum mechanics calculations are performed on the generated adversarial conformations to obtain their forces, and these data will be used for further training.

[0141] Step S4: Use adversarial molecules to perform adversarial training on the deep force field model. This step is the core of adversarial training. The model is trained by adversarial molecule samples to enhance the model's predictive ability and stability for extreme situations.

[0142] Step S5: Determine whether the model has converged. This step is a checkpoint in the training process to confirm whether the model has achieved the expected stability and accuracy.

[0143] Step S6: Obtaining a simulation-stable model. Once the model converges, what is finally obtained is a deep force field model that can stably predict the molecular dynamics state.

[0144] In summary, in the embodiments of the present application, the deep force field model is trained by the first loss function to ensure the basic ability of the model to predict molecular properties and dynamic states, and then the molecular data set is input into the pre-trained model and the prediction of the dynamic state is obtained through the adversarial sample generation algorithm, so as to further introduce adversarial conformations and non-extremeize the stress conditions in extreme cases, thereby ensuring the stability and reliability of the prediction results. Finally, the data set containing the adversarial conformation is trained by the second loss function, so that the deep force field model finally obtained can remain stable when predicting extreme forces, thereby avoiding the collapse of the simulation process. Therefore, the method based on the embodiments of the present application solves the problem in the related art that when performing molecular dynamics simulations through a force field model based on machine learning, extreme stress conditions will be predicted, thereby causing the deep mechanics model to be unable to achieve continuous and stable molecular dynamics simulations.

[0145] like Figure 6 As shown, the embodiment of the present application also discloses an adversarial training device 40 for a deep force field model for stable simulation, comprising:

[0146] The initial training set module 401 is used to establish a molecular data set; the molecular data set includes the original conformation of each molecule in a plurality of molecules; the original conformation is used to characterize the actual dynamic state of the molecule;

[0147] A pre-training module 402 is used to use the molecular data set as first training data, and train the initial model through the first training data and a first loss function to obtain a pre-training model;

[0148] The adversarial conformation set module 403 is used to input the molecular data set into the pre-trained model to obtain the predicted dynamic state of the molecules in the molecular data set, and expand the molecular data set into an adversarial conformation data set according to the predicted dynamic state by using the adversarial sample generation algorithm; the adversarial conformation data set includes the original conformation of the molecule and the adversarial conformation; the adversarial conformation is used to characterize the dynamic state of the molecule under mechanical perturbation;

[0149] Among them, the adversarial sample generation algorithm determines the adversarial conformation of the molecule through the following formula:

[0150] ;

[0151] in, Represents the updated spatial coordinate value of the molecule; r Represents the actual spatial coordinate value of the molecule in the molecular data set; represents the identity matrix; Represents the preset disturbance coefficient, which is a scalar used to represent the magnitude of the disturbance; represents a symbolic function; Indicates the gradient of the spatial coordinate value; J 1 represents the first loss function; θ represents the preset weight parameter; X represents a molecule; y Represents molecules X The actual dynamic state of

[0152] The final training module 404 is used to use the adversarial conformation data set as the second training data, and train the pre-trained model through the second training data and the second loss function to obtain a deep force field model.

[0153] Optionally, the pre-training module 402 includes:

[0154] A first prediction submodule is used to input the molecular data set into the deep force field model to be trained to obtain the predicted dynamic states of the molecules in the molecular data set;

[0155] A first loss submodule, for establishing a first loss function according to the predicted kinetic states of the molecules in the molecular data set and the actual kinetic states;

[0156] The first training submodule is used to train the deep force field model to be trained by using a first loss function, and obtain a pre-trained model when the output value of the deep force field model converges.

[0157] Optionally, the confrontation conformation set module 403 includes:

[0158] A second prediction submodule is used to input the molecular data set into the pre-trained model to obtain the predicted kinetic states of the molecules in the molecular data set;

[0159] The adversarial generation submodule is used to use the predicted kinetic state as the input value of the adversarial sample generation algorithm to calculate the adversarial conformation of the molecule.

[0160] Optionally, the dynamic state includes spatial coordinate values ​​and force values, and the antagonism generation submodule includes:

[0161] A loss acquisition unit, used to acquire a first loss function when the pre-training model training is completed;

[0162] A coordinate updating unit, used for determining an increment of the spatial coordinate value according to a gradient of the spatial coordinate value of the first loss function to update the spatial coordinate value of the molecule;

[0163] The force updating unit is used to determine the force value of the molecule after the updated spatial coordinate value according to the updated spatial coordinate value, and determine the dynamic state of the molecule after the updated spatial coordinate value and the force value as the antagonistic conformation of the molecule.

[0164] Optionally, the dynamic state includes a force value and an energy value, and the first loss submodule includes:

[0165] A first expectation unit is used to establish a molecular force expectation value of the molecular data set according to the predicted force value of the molecule in the molecular data set and the actual force value, and to establish a molecular energy expectation value of the molecular data set according to the predicted energy value of the molecule in the molecular data set and the actual energy value; the molecular force expectation value is the expectation of the square of the second-order norm of the difference between the predicted force value of the molecule in the molecular data set and the actual force value, and the molecular energy expectation value is the expectation of the square of the second-order norm of the difference between the predicted energy value of the molecule in the molecular data set and the actual energy value;

[0166] The first loss unit is used to take the sum of the expected value of the molecular energy and the expected value of the molecular force as the first loss function.

[0167] Optionally, the final training module 404 includes:

[0168] The second prediction submodule is used to input the adversarial conformational dataset into the pre-trained model to obtain the predicted dynamic states of the molecules in the adversarial conformational dataset;

[0169] A second loss submodule, for establishing a second loss function according to the predicted dynamic states of the molecules in the adversarial conformational dataset and the actual dynamic states;

[0170] The second training submodule is used to train the pre-trained model through a second loss function, and obtain a trained deep force field model when the output value of the pre-trained model converges.

[0171] Optionally, the dynamic state includes a force value, and the second loss submodule includes:

[0172] The second loss unit is used to establish the expected value of the antagonistic force of the antagonistic conformational data set based on the predicted force value of the molecule in the antagonistic conformational data set and the actual force value, and use the expected value of the antagonistic force as the second loss function; the expected value of the antagonistic force is the expectation of the square of the second-order norm of the difference between the predicted force value of the molecule in the antagonistic conformational data set and the actual force value.

[0173] like Figure 7 As shown, the embodiment of the present application also discloses a molecular dynamics simulation device 50, comprising:

[0174] The data preparation module 501 is used to obtain an experimental data set to be simulated; the experimental data set includes multiple original conformations of molecules; the original conformations are used to characterize the actual dynamic state of the molecules;

[0175] The data simulation module 502 is used to input the experimental data set into the deep force field model trained by the adversarial training method of the deep force field model for stable simulation disclosed in the embodiments of the present application, so as to obtain the predicted dynamic state of the molecules in the experimental data set as the simulation result of molecular dynamics.

[0176] In summary, in the embodiments of the present application, the deep force field model is trained by the first loss function to ensure the basic ability of the model to predict molecular properties and dynamic states, and then the molecular data set is input into the pre-trained model and the prediction of the dynamic state is obtained through the adversarial sample generation algorithm, so as to further introduce adversarial conformations and non-extremeize the stress conditions in extreme cases, thereby ensuring the stability and reliability of the prediction results. Finally, the data set containing the adversarial conformation is trained by the second loss function, so that the deep force field model finally obtained can remain stable when predicting extreme forces, thereby avoiding the collapse of the simulation process. Therefore, the method based on the embodiments of the present application solves the problem in the related art that when performing molecular dynamics simulations through a force field model based on machine learning, extreme stress conditions will be predicted, thereby causing the deep mechanics model to be unable to achieve continuous and stable molecular dynamics simulations.

[0177] The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, each process of the above-mentioned adversarial training method for the deep force field model for stable simulation and the molecular dynamics simulation method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. Among them, the computer-readable storage medium, such as read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0178] Figure 8 700 is a block diagram of an electronic device 700 provided in an embodiment of the present application. For example, the electronic device 700 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0179] Reference Figure 8 , the electronic device 700 may include one or more of the following components: a processing component 702 , a memory 704 , a power component 706 , a multimedia component 708 , an audio component 710 , an input / output (I / O) interface 712 , a sensor component 714 , and a communication component 716 .

[0180] The processing component 702 generally controls the overall operation of the electronic device 700, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 702 may include one or more processors 720 to execute instructions to complete all or part of the steps of the above-mentioned adversarial training method for the deep force field model for stable simulation and the molecular dynamics simulation method. In addition, the processing component 702 may include one or more modules to facilitate the interaction between the processing component 702 and other components. For example, the processing component 702 may include a multimedia module to facilitate the interaction between the multimedia component 708 and the processing component 702.

[0181] The memory 704 is used to store various types of data to support the operation of the electronic device 700. Examples of such data include instructions for any application or method operating on the electronic device 700, contact data, phone book data, messages, pictures, multimedia, etc. The memory 704 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0182] The power supply component 706 provides power to the various components of the electronic device 700. The power supply component 706 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the electronic device 700.

[0183] The multimedia component 708 includes a screen that provides an output interface between the electronic device 700 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the demarcation of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 708 includes a front camera and / or a rear camera. When the electronic device 700 is in an operating mode, such as a shooting mode or a multimedia mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.

[0184] The audio component 710 is used to output and / or input audio signals. For example, the audio component 710 includes a microphone (MIC), and when the electronic device 700 is in an operation mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is used to receive an external audio signal. The received audio signal can be further stored in the memory 704 or sent via the communication component 716. In some embodiments, the audio component 710 also includes a speaker for outputting audio signals.

[0185] I / O interface 712 provides an interface between processing component 702 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include but are not limited to: home button, volume button, start button, and lock button.

[0186] The sensor assembly 714 includes one or more sensors for providing various aspects of status assessment for the electronic device 700. For example, the sensor assembly 714 can detect the open / closed state of the electronic device 700, the relative positioning of components, such as the display and keypad of the electronic device 700, and the sensor assembly 714 can also detect the position change of the electronic device 700 or a component of the electronic device 700, the presence or absence of user contact with the electronic device 700, the orientation or acceleration / deceleration of the electronic device 700, and the temperature change of the electronic device 700. The sensor assembly 714 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor assembly 714 may also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 714 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0187] The communication component 716 is used to facilitate wired or wireless communication between the electronic device 700 and other devices. The electronic device 700 can access a wireless network based on a communication standard, such as WiFi, a carrier network (such as 2G, 3G, 4G or 7G), or a combination thereof. In an exemplary embodiment, the communication component 716 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 716 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0188] In an exemplary embodiment, the electronic device 700 can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to implement the adversarial training method for the deep force field model for stable simulation and the molecular dynamics simulation method provided in the embodiments of the present application.

[0189] In an exemplary embodiment, a non-temporary computer-readable storage medium including instructions is also provided, such as a memory 704 including instructions, and the instructions can be executed by a processor 720 of an electronic device 700 to complete the above-mentioned adversarial training method for a deep force field model for stable simulation and a molecular dynamics simulation method. For example, the non-temporary storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a tape, a floppy disk, an optical data storage device, etc.

[0190] Fig. 9 8 is a block diagram of an electronic device 800 according to an exemplary embodiment. For example, the electronic device 800 may be provided as a server. Fig. 9 The electronic device 800 includes a processing component 822, which further includes one or more processors, and a memory resource represented by a memory 832 for storing instructions that can be executed by the processing component 822, such as an application. The application stored in the memory 832 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 822 is configured to execute instructions to execute the adversarial training method and molecular dynamics simulation method of the deep force field model for stable simulation provided in an embodiment of the present application.

[0191] The electronic device 800 may also include a power supply component 826 configured to perform power management of the electronic device 800, a wired or wireless network interface 850 configured to connect the electronic device 800 to a network, and an input / output (I / O) interface 858. The electronic device 800 may operate based on an operating system stored in the memory 832, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™ or the like.

[0192] An embodiment of the present application also provides a computer program product, including a computer program, which implements an adversarial training method and a molecular dynamics simulation method for a deep force field model for stable simulation when the computer program is executed by a processor.

[0193] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the application disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are only considered as exemplary, and the true scope and spirit of the present application are indicated by the claims of the present application.

[0194] It will be understood that the present application is not limited to the exact construction that has been described and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

[0195] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0196] It is easy for a person skilled in the art to think that any combination of the above embodiments is feasible, so any combination of the above embodiments is an implementation scheme of the present application. However, due to space limitations, this specification will not describe them in detail here.

[0197] The adversarial training method and the molecular dynamics simulation method for the deep force field model for stable simulation provided herein are not inherently related to any specific computer, virtual system or other device. Various general systems can also be used together with the teaching based on this. According to the above description, it is obvious to construct the structure required for the system with the scheme of the present application. In addition, the present application is not directed to any specific programming language. It should be understood that the content of the present application described herein can be implemented using various programming languages, and the above description of the specific language is to disclose the best implementation mode of the present application.

[0198] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0199] Similarly, it should be understood that in order to streamline the present application and aid in understanding one or more of the various application aspects, in the above description of the exemplary embodiments of the present application, the various features of the present application are sometimes grouped together into a single embodiment, figure, or description thereof. However, the disclosed method should not be interpreted as reflecting the following intention: the claimed application requires more features than the features explicitly recited in each claim. More specifically, as reflected in the claims, the application aspects lie in less than all the features of the single embodiment disclosed above. Therefore, the claims that follow the specific embodiment are hereby expressly incorporated into the specific embodiment, with each claim itself serving as a separate embodiment of the present application.

[0200] Those skilled in the art will appreciate that the modules in the equipment in the embodiment can be adaptively changed and arranged in one or more equipment different from the embodiment. The modules or units or components in the embodiment can be combined into a module or unit or component, and in addition they can be divided into multiple submodules or subunits or subassemblies. Except that at least some of such features and / or processes or units are mutually exclusive, all features disclosed in the application and all processes or units of any method or equipment disclosed in this manner can be combined in any combination. Unless otherwise clearly stated, each feature disclosed in the application can be replaced by alternative features that provide the same, equivalent or similar purpose.

[0201] In addition, those skilled in the art will appreciate that, although some embodiments described herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present application and form different embodiments. For example, in the present application, any one of the claimed embodiments may be used in any combination.

[0202] The various component embodiments of the present application may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or a digital signal processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components of the adversarial training method for a deep force field model for stable simulation and the molecular dynamics simulation method according to the embodiment of the present application. The present application may also be implemented as a device or apparatus program (e.g., a computer program and a computer program product) for executing part or all of the adversarial training method for a deep force field model for stable simulation and the molecular dynamics simulation method described herein. Such a program implementing the present application may be stored on a computer-readable medium, or may be in the form of one or more signals. Such a signal may be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0203] It should be noted that the above embodiments illustrate the present application rather than limit the present application, and that those skilled in the art may design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets should not be constructed as a limitation to the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of multiple such elements. The present application may be implemented by means of hardware including several different elements and by means of a suitably programmed computer. In a unit claim that lists several devices, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order. These words may be interpreted as names.

Claims

1. An adversarial training method for a deep force field model for stable simulation, characterized in that: include: Create molecular datasets; The molecular dataset comprises an original conformation of each of the molecules in a plurality of molecules; The original conformation is used to characterize the actual dynamic state of the molecule; Using the molecular data set as first training data, and training the initial model using the first training data and a first loss function to obtain a pre-trained model; Inputting the molecular data set into the pre-trained model to obtain the predicted kinetic states of the molecules in the molecular data set, and expanding the molecular data set into an adversarial conformational data set using an adversarial sample generation algorithm according to the predicted kinetic states; the adversarial conformational data set includes the original conformation of the molecule and the adversarial conformation; The antagonistic conformation is used to characterize the dynamic state of the molecule under mechanical perturbation; Using the adversarial conformation data set as second training data, and training the pre-trained model using the second training data and a second loss function to obtain a deep force field model; The adversarial sample generation algorithm determines the adversarial conformation of the molecule through the following formula: ; in, Represents the updated spatial coordinate value of the molecule; r Represents the actual spatial coordinate value of the molecule in the molecular data set; represents the identity matrix; Represents the preset disturbance coefficient, which is a scalar used to represent the magnitude of the disturbance; represents a symbolic function; Indicates the gradient of the spatial coordinate value; J 1 represents the first loss function; θ represents the preset weight parameter; X represents a molecule; y Represents molecules X The actual dynamic state.

2. The adversarial training method for a deep force field model for stable simulation according to claim 1, characterized in that: The method of using the molecular data set as first training data and training the initial model using the first training data and a first loss function to obtain a pre-trained model includes: Inputting the molecular data set into the deep force field model to be trained to obtain predicted dynamic states of the molecules in the molecular data set; Establishing the first loss function according to the predicted kinetic states of the molecules in the molecular data set and the actual kinetic states; Training the deep force field model to be trained by using the first loss function, and obtaining the pre-trained model when the output value of the deep force field model converges; The first loss function is: ; in, represents the first loss function, E e represents the expected value of molecular energy, E f represents the expected value of the molecular force, represents the force value predicted by the molecule in the molecular dataset, E({r}) Indicates the actual force value of the molecule in the molecular data set, represents the energy value of the molecule prediction in the molecule dataset, F Represents the actual energy value of the molecule in the molecular data set, To find the sign of the second-order norm.

3. The adversarial training method for a deep force field model for stable simulation according to claim 1, characterized in that: The step of inputting the molecular dataset into the pre-trained model to obtain a predicted kinetic state of the molecules in the molecular dataset, and expanding the molecular dataset into an adversarial conformational dataset using an adversarial sample generation algorithm according to the predicted kinetic state, comprises: Inputting the molecular data set into the pre-trained model to obtain predicted kinetic states of the molecules in the molecular data set; Using the predicted kinetic state as an input value of the adversarial sample generation algorithm to calculate the adversarial conformation of the molecule; The predicted kinetic state is used as an input value of the adversarial sample generation algorithm to calculate the adversarial conformation of the molecule, comprising: Obtaining the first loss function when the pre-training model training is completed; Determine the increment of the spatial coordinate value according to the gradient of the first loss function to the spatial coordinate value, so as to update the spatial coordinate value of the molecule; The force value of the molecule after the spatial coordinate value is updated is determined according to the updated spatial coordinate value, and the dynamic state of the molecule after the spatial coordinate value and the force value are updated is determined as the antagonistic conformation of the molecule.

4. The adversarial training method for a deep force field model for stable simulation according to claim 1, characterized in that: The method further comprises: using the adversarial conformation dataset as second training data, and training the pre-trained model by using the second training data and a second loss function to obtain a deep force field model, including: Inputting the adversarial conformational dataset into the pre-trained model to obtain the predicted dynamic states of the molecules in the adversarial conformational dataset; Establishing the second loss function according to the predicted dynamic state of the molecule in the adversarial conformational dataset and the actual dynamic state; The pre-trained model is trained by using the second loss function, and a trained deep force field model is obtained when the output value of the pre-trained model converges; The second loss function is: ; in, represents the second loss function, represents the expected value of the resisting force, represents the force value predicted by the molecule in the confrontation conformation dataset, F’ represents the actual force value of the molecule in the confrontation conformation dataset, To find the sign of the second-order norm.

5. A molecular dynamics simulation method, characterized in that: include: Acquire an experimental data set to be simulated; the experimental data set includes a plurality of original conformations of the molecule; the original conformations are used to characterize the actual dynamic state of the molecule; The experimental data set is input into a deep force field model trained by the adversarial training method of a deep force field model for stable simulation as described in any one of claims 1 to 4 to obtain the predicted dynamic state of the molecules in the experimental data set as a simulation result of molecular dynamics.

6. An adversarial training device for a deep force field model for stable simulation, characterized in that: include: Initial training set module, used to build molecular data sets; The molecular dataset comprises an original conformation of each of the molecules in a plurality of molecules; The original conformation is used to characterize the actual dynamic state of the molecule; A pre-training module, used to use the molecular data set as first training data, and train the initial model through the first training data and a first loss function to obtain a pre-training model; An adversarial conformation set module, used for inputting the molecular data set into the pre-trained model to obtain the predicted dynamic state of the molecules in the molecular data set, and expanding the molecular data set into an adversarial conformation data set according to the predicted dynamic state by using an adversarial sample generation algorithm; the adversarial conformation data set includes the original conformation of the molecule and the adversarial conformation; the adversarial conformation is used to characterize the dynamic state of the molecule under mechanical perturbation; A final training module, used for taking the adversarial conformation data set as second training data, and training the pre-trained model by using the second training data and a second loss function to obtain a deep force field model; The adversarial sample generation algorithm determines the adversarial conformation of the molecule through the following formula: ; in, Represents the updated spatial coordinate value of the molecule; r Represents the actual spatial coordinate value of the molecule in the molecular data set; represents the identity matrix; Represents the preset disturbance coefficient, which is a scalar used to represent the magnitude of the disturbance; represents a symbolic function; Indicates the gradient of the spatial coordinate value; J 1 represents the first loss function; θ represents the preset weight parameter; X represents a molecule; y Represents molecules X The actual dynamic state.

7. A molecular dynamics simulation device, characterized in that: include: Data preparation module, used to obtain the experimental data set to be simulated; The experimental data set includes a plurality of original conformations of the molecule; the original conformations are used to characterize the actual dynamic state of the molecule; A data simulation module is used to input the experimental data set into a deep force field model trained by the adversarial training method of the deep force field model for stable simulation as described in any one of claims 1 to 4, so as to obtain the predicted dynamic state of the molecules in the experimental data set as a simulation result of molecular dynamics.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

9. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the method according to any one of claims 1 to 5 when executed by the processor.

10. A computer program product, characterized in that The computer program product stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Citation Information

Patent Citations

  • Molecular docking scoring method based on deep learning and inter-atomic acting force weighting

    CN116612832A

  • Molecular conformation generation method based on conditional generative adversarial network and storage medium

    CN116646027A

  • Molecular representation learning-oriented hierarchical geometric graph neural network construction method and device

    CN117952144A

  • Method and device for predicting properties of metal material based on machine learning force field and medium

    CN118136186A