Molecular dynamics simulation prediction method, device, equipment and storage medium

Abstract

Embodiments of the present application provide a molecular dynamics simulation prediction method, device, equipment and storage medium, and relate to the technical field of data processing. A protein sequence to be tested corresponding to a protein is obtained, the protein sequence to be tested is input into a graph network structure for feature extraction to obtain graph feature data corresponding to each peptide unit, the graph feature data is input into a gating structure for feature processing to obtain an allocation result corresponding to each expert network, and the graph feature data and the allocation result are input into at least one expert model for data prediction to obtain a predicted potential energy and a predicted force field corresponding to each peptide unit. Heterogeneous reasoning is performed, the graph network structure is deployed in GPU display memory, efficient data reading and feature extraction are performed, the hybrid expert model is deployed in CPU memory, parallel reasoning processes are performed by using CPU cores, and the reasoning speed of the hybrid expert model is accelerated by using the processing advantages of a heterogeneous cluster. Moreover, the size of the hybrid expert model can be increased in the memory, and the reasoning accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to molecular dynamics simulation and prediction methods, apparatus, devices, and storage media. Background Technology

[0002] Mixture of Experts (MoE) models integrate multiple expert models to construct a novel model. Compared to dense fully connected models, this model possesses stronger representational capabilities and is widely used in numerous fields such as molecular dynamics simulations, weather forecasting, and protein structure prediction. However, because MoE models are typically loaded into GPU memory and consume a significant amount of redundant GPU memory during inference, the scale of runnable MoE models is limited by GPU memory and computing resources. A small scale makes it difficult to fully leverage the advantages of MoE models, thus limiting computational accuracy; conversely, increasing the scale to improve accuracy can significantly reduce computational speed due to insufficient resources.

[0003] In related technologies, to simultaneously improve inference accuracy and speed, methods such as optimizing gating mechanisms and sparsifying hybrid expert models are employed to enhance their training and inference speed. However, this approach still limits the size of the hybrid expert model, making it difficult to achieve a good balance between inference accuracy and speed. Summary of the Invention

[0004] The main objective of this application is to propose a molecular dynamics simulation prediction method, apparatus, device, and storage medium, while improving the inference accuracy and speed of hybrid expert models.

[0005] To achieve the above objectives, a first aspect of this application proposes a molecular dynamics simulation prediction method, executed by a pre-trained dynamics simulation prediction model. The dynamics simulation prediction model includes a graph network structure and a gating structure deployed on a GPU device, and at least one parallel expert model deployed on a CPU device. The method includes:

[0006] Obtain the sequence of the protein to be tested corresponding to the protein, wherein the sequence of the protein to be tested includes multiple peptide units;

[0007] The protein sequence to be tested is input into the graph network structure for feature extraction to obtain graph feature data corresponding to each peptide unit. The graph feature data is then input into the gating structure for feature processing to obtain the allocation result corresponding to each expert network.

[0008] The graph feature data and the allocation result are input into the at least one expert model for data prediction to obtain the predicted potential energy and predicted force field corresponding to each peptide unit.

[0009] In one embodiment, obtaining the protein sequence to be tested includes:

[0010] Obtain the initial protein data corresponding to the protein, and divide the initial protein data with peptide bonds as the center to obtain multiple peptide units;

[0011] The three-dimensional coordinates of carbon atoms, oxygen atoms, nitrogen atoms, and residues in each peptide unit are obtained as the three-dimensional coordinates of the peptide unit, thus obtaining the protein sequence to be tested.

[0012] In one embodiment, the graph network structure includes a graph convolutional network and a linear transformation network. The step of inputting the protein sequence to be tested into the graph network structure for feature extraction to obtain graph feature data corresponding to each peptide unit includes:

[0013] The graph convolutional network is used to perform graph convolution operations on each peptide unit in the protein sequence to be tested to obtain the first feature data.

[0014] The first feature data is linearly transformed using the linear transformation network to obtain the graph feature data corresponding to the peptide unit.

[0015] In one embodiment, the step of using the linear transformation network to perform a linear transformation on the first feature data to obtain the graph feature data corresponding to the peptide unit includes:

[0016] Perform a convolution operation on the first feature data to obtain multiple convolutional feature data;

[0017] The multiple convolutional feature data are concatenated and then split into query data, key data, and value data;

[0018] The query data and the key data are respectively processed with the activation function to obtain intermediate query data and intermediate key data;

[0019] The intermediate query data, the intermediate key data, and the value data are multiplied by a matrix, and then a convolution operation is performed to obtain the graph feature data.

[0020] In one embodiment, the step of inputting the graph feature data and the allocation result into the at least one expert model for data prediction to obtain the predicted potential energy and predicted force field corresponding to the peptide unit includes:

[0021] Based on the allocation result, at least one graph feature data corresponding to each expert model is obtained as representation data;

[0022] The expert model is used to predict the characterization data to obtain the simulated trajectory position corresponding to the three-dimensional coordinates of the peptide unit, and the predicted potential energy and predicted force field of the peptide unit are obtained based on the simulated trajectory position.

[0023] Summarize the predicted potential energy and predicted force field outputs from all the expert models.

[0024] In one embodiment, the step of using the expert model to predict the characterization data and obtain the simulated trajectory position corresponding to the three-dimensional coordinates of the peptide unit includes:

[0025] Based on the initial velocity information of the protein, velocity and acceleration data of each peptide unit at different time steps are obtained;

[0026] The three-dimensional coordinates of the peptide unit corresponding to each time step are updated at least using the velocity data and the acceleration data;

[0027] After a preset number of time steps of dynamic simulation, the three-dimensional coordinates of the peptide unit corresponding to the last time step are obtained as the simulated trajectory position.

[0028] In one embodiment, the training process of the dynamic simulation prediction model includes the following steps:

[0029] Construct a sample dataset, which includes multiple training samples, each training sample including training protein sequences and training labels, the training labels including potential energy labels and force field labels;

[0030] The training samples are input into the dynamic simulation prediction model for data processing to obtain the training prediction potential energy and the training prediction force field.

[0031] The potential energy loss value is calculated based on the potential energy label and the training predicted potential energy; the force field loss value is calculated based on the force field label and the training predicted force field; and the total loss value is calculated based on the potential energy loss value and the force field loss value.

[0032] The model parameters of the dynamic simulation prediction model are adjusted according to the total loss value until the training termination condition is met, thus obtaining the trained dynamic simulation prediction model.

[0033] To achieve the above objectives, a second aspect of this application proposes a molecular dynamics simulation prediction device, executed by a pre-trained dynamics simulation prediction model. The dynamics simulation prediction model includes a graph network structure and a gating structure deployed on a GPU device, and at least one parallel expert model deployed on a CPU device. The device comprises:

[0034] Data acquisition module: used to acquire the sequence of the protein to be tested, which includes multiple peptide units;

[0035] Feature extraction module: used to input the protein sequence to be tested into the graph network structure for feature extraction, to obtain graph feature data corresponding to each peptide unit, and to input the graph feature data into the gating structure for feature processing, to obtain the allocation result corresponding to each expert network;

[0036] Expert prediction module: used to input the graph feature data and the allocation result into the at least one expert model to perform data prediction, and obtain the predicted potential energy and predicted force field corresponding to each peptide unit.

[0037] To achieve the above objectives, a third aspect of the present application provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described in the first aspect.

[0038] To achieve the above objectives, a fourth aspect of the present application provides a storage medium that stores a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0039] The molecular dynamics simulation prediction method, apparatus, device, and storage medium proposed in this application are executed by a pre-trained dynamics simulation prediction model. This model includes a graph network structure and a gating structure deployed on a GPU device, and at least one parallel expert model deployed on a CPU device. The method involves acquiring the protein sequence to be tested, which includes multiple peptide units. The protein sequence is input into the graph network structure for feature extraction, yielding graph feature data for each peptide unit. This graph feature data is then input into the gating structure for feature processing to obtain the allocation result for each expert network. Finally, the graph feature data and allocation results are input into at least one expert model for data prediction, yielding the predicted potential energy and predicted force field for each peptide unit. In this application, the dynamics simulation prediction model undergoes heterogeneous inference. The graph network structure, primarily used for feature extraction, is deployed in GPU memory for efficient data reading and feature extraction. The hybrid expert model is deployed in CPU memory, utilizing multiple CPU cores for parallel inference. This fully leverages the processing advantages of different computing resources in a heterogeneous cluster, accelerating the inference speed of the hybrid expert model. In addition, since hybrid expert models are deployed in memory, they can be scaled up and their inference accuracy improved. Attached Figure Description

[0040] Figure 1This is a schematic diagram of the deployment framework of the dynamic simulation prediction model provided in the embodiments of this application.

[0041] Figure 2 This is a flowchart of the molecular dynamics simulation and prediction method provided in the embodiments of this application.

[0042] Figure 3 This is a flowchart of obtaining the sequence of the protein to be tested provided in the embodiments of this application.

[0043] Figure 4 This is a flowchart provided in an embodiment of the present application, in which the sequence of the protein to be tested is input into a graph network structure for feature extraction to obtain graph feature data.

[0044] Figure 5 This is a schematic diagram of the graph network structure provided in the embodiments of this application.

[0045] Figure 6 This is a flowchart provided in an embodiment of the present application, which uses a linear transformation network to perform a linear transformation on the first feature data to obtain the graph feature data corresponding to the peptide unit.

[0046] Figure 7 This is another schematic diagram of the dynamic simulation prediction model provided in the embodiments of this application.

[0047] Figure 8 This is a flowchart provided in this application embodiment, in which graph feature data and allocation results are input into at least one expert model for data prediction to obtain the predicted potential energy and predicted force field corresponding to the peptide unit.

[0048] Figure 9 This is a flowchart provided in an embodiment of the present application, which uses an expert model to predict the data of the characterization data and obtain the simulated trajectory position corresponding to the three-dimensional coordinates of the peptide unit.

[0049] Figure 10 This is a schematic diagram of the training process of the dynamic simulation prediction model provided in the embodiments of this application.

[0050] Figure 11 This is a structural block diagram of a molecular dynamics simulation and prediction device provided in another embodiment of this application.

[0051] Figure 12 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0055] First, let's analyze some of the terms used in this application:

[0056] Artificial Intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0057] Mixture of Experts (MoE) models integrate multiple expert models to construct a novel model. Compared to dense fully connected models, this model possesses stronger representational capabilities and is widely used in numerous fields such as molecular dynamics simulations, weather forecasting, and protein structure prediction. However, because MoE models are typically loaded into GPU memory and consume a significant amount of redundant GPU memory during inference, the scale of runnable MoE models is limited by GPU memory and computing resources. A small scale makes it difficult to fully leverage the advantages of MoE models, thus limiting computational accuracy; conversely, increasing the scale to improve accuracy can significantly reduce computational speed due to insufficient resources.

[0058] In related technologies, to simultaneously improve inference accuracy and speed, methods such as optimizing gating mechanisms and sparsifying hybrid expert models are employed to enhance their training and inference speed. However, this approach still limits the size of the hybrid expert model, making it difficult to achieve a good balance between inference accuracy and speed.

[0059] Based on this, embodiments of this application provide a molecular dynamics simulation prediction method, apparatus, device, and storage medium. The method performs heterogeneous inference on the dynamics simulation prediction model. A graph network structure, primarily used for feature extraction, is deployed in GPU memory for efficient data reading and feature extraction. A hybrid expert model is deployed in CPU memory, utilizing multiple CPU cores for parallel inference. This fully leverages the processing advantages of different computing resources in a heterogeneous cluster, accelerating the inference speed of the hybrid expert model. Furthermore, because the hybrid expert model is deployed in memory, its scale can be increased, improving inference accuracy.

[0060] This application provides a molecular dynamics simulation and prediction method, apparatus, device, and storage medium, which are specifically described through the following embodiments. First, the molecular dynamics simulation and prediction method in this application embodiment is described.

[0061] This application's embodiments can acquire and process relevant data based on artificial intelligence (AI) technology. AI is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine that can react in a way similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0062] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0063] The molecular dynamics simulation and prediction method provided in this application relates to the field of data processing technology. This method can be applied to a terminal, a server, or a computer program running on either a terminal or a server. For example, the computer program can be a native program or software module in an operating system; it can be a native application (APP), i.e., a program that needs to be installed in the operating system to run, such as a client supporting molecular dynamics simulation and prediction, i.e., a program that only needs to be downloaded to a browser environment to run; or it can be a small program that can be embedded into any APP. In short, the above-mentioned computer program can be any form of application, module, or plugin. The terminal communicates with the server via a network. The molecular dynamics simulation and prediction method can be executed by the terminal or the server, or by the terminal and the server working together.

[0064] In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc. The server can be a standalone server, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms; it can also be a service node in a blockchain system, where the service nodes form a peer-to-peer (P2P) network. The P2P protocol is an application layer protocol running on top of the Transmission Control Protocol (TCP). The terminal and server can connect via Bluetooth, Universal Serial Bus (USB), or a network, etc., and this embodiment does not impose any limitations.

[0065] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0066] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0067] The dynamic simulation prediction model in the embodiments of this application is described below.

[0068] In one embodiment, reference is made to Figure 1 , Figure 1 This is a schematic diagram of the deployment framework of the dynamic simulation prediction model provided in the embodiments of this application.

[0069] Figure 1 The dynamic simulation prediction model includes graph network structures and gating structures deployed on GPU devices and hybrid expert models deployed on CPU devices. The hybrid expert model includes at least one parallel expert model, such as Expert1, Expert2, ..., ExpertN in the figure.

[0070] Specifically, data processing is performed on the GPU using graph network and gating structures. The output data is then transferred to memory via a high-speed serial computer extended bus standard (Peripheral Component Interconnect Express, PCIe) interface. The PCIe interface enables data communication between the GPU and memory, ensuring data transfer speed and stability. Next, the data in memory is fed into a hybrid expert model on the CPU for prediction, yielding the prediction results.

[0071] This application's embodiments utilize heterogeneous inference with a dynamic simulation prediction model. The graph network structure, primarily used for feature extraction, is deployed in GPU memory for efficient data reading and feature extraction. The hybrid expert model is deployed in CPU memory, leveraging multiple CPU cores for parallel inference. This fully utilizes the processing advantages of different computing resources in a heterogeneous cluster, accelerating the inference speed of the hybrid expert model. Furthermore, because the hybrid expert model is deployed in memory, its scale can be increased, improving inference accuracy.

[0072] The molecular dynamics simulation prediction method in the embodiments of this application is described below in conjunction with the deployment framework of the dynamics simulation prediction model.

[0073] Figure 2 This is an optional flowchart of the molecular dynamics simulation and prediction method provided in the embodiments of this application. Figure 2 The method may include, but is not limited to, steps 110 to 130. It is also understood that this embodiment... Figure 2 The order of steps 110 to 130 is not specifically limited. The order of steps can be adjusted or some steps can be reduced or added according to actual needs.

[0074] Step 110: Obtain the sequence of the protein to be tested corresponding to the protein.

[0075] In one embodiment, the protein is a peptide chain composed of multiple amino acids arranged in a "dehydration condensation" manner, and the peptide chain is dynamically folded to obtain a 3D structure. The protein to be tested is referred to as a test protein sequence, which includes multiple peptide units.

[0076] In one embodiment, reference is made to Figure 3 , Figure 3 This is a flowchart of obtaining the sequence of a protein to be tested provided in an embodiment of this application, which specifically includes the following steps:

[0077] Step 310: Obtain the initial protein data corresponding to the protein, and divide the initial protein data into multiple peptide units centered on peptide bonds.

[0078] In one embodiment, carbon atoms (C), oxygen atoms (O), and nitrogen atoms (N) in the protein form peptide bonds (-CO-NH-). Peptide bonds are used to link amino acid residues into a peptide chain. Amino acid residues are the parts remaining after the removal of water molecules from amino acids. Specifically, the carbonyl carbon atom (C) is linked to the oxygen atom (O) by a double bond, and this carbon atom is also linked to the nitrogen atom (N) by a single bond. The nitrogen atom is then linked to a carbon atom in another amino acid residue. In other words, multiple amino acid residues in the protein are linked by peptide bonds. Therefore, in this embodiment, the initial protein data is segmented according to peptide bonds to obtain peptide units associated with the number of residues. Assuming a protein sequence is 40 units long, which can be understood as containing 40 residues, cutting each residue and its corresponding peptide bond as a single peptide unit yields 39 peptide units.

[0079] Step 320: Obtain the three-dimensional coordinates of carbon atoms, oxygen atoms, nitrogen atoms, and residues in each peptide unit as the three-dimensional coordinates of the peptide unit to obtain the protein sequence to be tested.

[0080] In one embodiment, since a peptide unit contains peptide bonds and residues, and a peptide bond contains nitrogen atoms, oxygen atoms and carbon atoms, the three-dimensional coordinates of the carbon atoms, oxygen atoms, nitrogen atoms and residues in each peptide unit are obtained as the three-dimensional coordinates of the peptide unit. After arranging them in order, the protein sequence to be tested is obtained.

[0081] In one embodiment, it is assumed that protein P i The corresponding initial protein data has a length of L, and the protein sequence to be tested is represented as follows:

[0082]

[0083] Where b represents the b-th peptide unit. These are the three-dimensional coordinates of carbon atoms at different positions. These are the three-dimensional coordinates of the oxygen atom. These are the three-dimensional coordinates of a nitrogen atom. This represents the three-dimensional coordinates of the residues linked to the peptide unit. The three-dimensional coordinates are the x, y, and z coordinate components in three-dimensional space.

[0084] Next, a kinetic simulation prediction model is used to predict the kinetics of the protein sequence to be tested.

[0085] Step 120: Input the protein sequence to be tested into the graph network structure for feature extraction to obtain the graph feature data corresponding to each peptide unit, and input the graph feature data into the gating structure for feature processing to obtain the allocation result corresponding to each expert network.

[0086] In one embodiment, combined with Figure 1 First, the protein sequence to be tested is fed into a graph network structure for feature extraction. The extracted data is then fed into a gate control structure. This process is described in detail below. (Refer to...) Figure 4 , Figure 4 This is a flowchart provided in this application embodiment for extracting features from a graph network structure by inputting the protein sequence to be tested, and obtaining graph feature data. The flowchart specifically includes the following steps:

[0087] Step 410: Use a graph convolutional network to perform graph convolution operations on each peptide unit in the protein sequence to be tested to obtain the first feature data.

[0088] In one embodiment, reference is made to Figure 5 , Figure 5 This is a schematic diagram of a graph network structure provided in an embodiment of this application. The graph network structure includes a first number of cascaded graph convolutional networks and a second number of cascaded linear transformation networks. As shown in the example, the first number can be 4, and the second number can be 6.

[0089] Reference Figure 5 The specific graph convolutional network includes cascaded graph convolutional layers, activation function layers, and normalization layers. Therefore, peptide units in the protein sequence to be tested are input one by one into the cascaded graph convolutional network. Through graph convolution operations, activation operations, and normalization operations, the first feature data corresponding to each peptide unit is finally obtained. It is understood that the activation function corresponding to the activation function layer in this embodiment can be set according to actual needs.

[0090] Step 420: Use a linear transformation network to perform a linear transformation on the first feature data to obtain the graph feature data corresponding to the peptide unit.

[0091] In one embodiment, reference is made to Figure 6 , Figure 6 This is a flowchart provided in this application embodiment of the process of using a linear transformation network to perform a linear transformation on the first feature data to obtain the graph feature data corresponding to the peptide unit, specifically including the following steps:

[0092] Step 610: Perform a convolution operation on the first feature data to obtain multiple convolutional feature data.

[0093] In one embodiment, reference is made to Figure 5 The linear transformation network includes multiple convolutional layers. The first convolutional layer performs a convolution operation on the first feature data to obtain initial convolutional data. The initial convolutional data is then fed into other convolutional layers in parallel for convolution processing to obtain multiple convolutional feature data.

[0094] Step 620: Concatenate multiple convolutional feature data and then split them into query data, key data, and value data.

[0095] In one embodiment, reference is made to Figure 5 Then, the convolutional feature data of all the convolutional layers are fed into the concatenation and splitting layer. After concatenating these convolutional feature data, they are split into three different vectors, namely query data q, key data k, and value data v.

[0096] Step 630: Apply the query data and key data to the activation function to obtain intermediate query data and intermediate key data respectively.

[0097] In one embodiment, reference is made to Figure 5 The query data is sent to the activation function layer to obtain intermediate query data, and the key data is sent to the activation function layer to obtain intermediate key data.

[0098] Step 640: Multiply the intermediate query data, intermediate key data, and value data by matrix multiplication, and then perform a convolution operation to obtain graph feature data.

[0099] In one embodiment, reference is made to Figure 5 After performing matrix multiplication on the intermediate key data and value data, the result is then multiplied by the intermediate query data to obtain the initial output data. The initial output data is then scaled to adjust the numerical range before being fed into the convolutional layer. The convolution result is then superimposed on the initial convolutional data to obtain the final graph feature data.

[0100] In one embodiment, reference is made to Figure 7 , Figure 7 This is another schematic diagram of the dynamic simulation prediction model provided in the embodiments of this application. Figure 7 The number of graph feature data obtained through the linear transformation network is related to the length L of the protein, for example, L = 500 as shown in the figure. That is to say, for the protein sequence to be tested, a corresponding graph feature data needs to be generated for each peptide unit. If the number of peptide units of a certain protein is insufficient, the corresponding default value is filled in as peptide units in the protein sequence to be tested. The default value can be an all-zero vector.

[0101] In one embodiment, reference is made to Figure 7After acquiring the graph feature data corresponding to each peptide unit, this graph feature data is input into a gating network for allocation, resulting in an allocation result for each expert network. This allocation result indicates the graph feature data that the corresponding expert network needs to process. The gating network distributes the graph feature data to the corresponding expert networks, so that different expert networks only need to process the allocated graph feature data. Once the gating network has completed the allocation of graph feature data to the corresponding expert networks, each expert network can simultaneously process one or more graph feature data it has received, thus achieving parallel processing. Since multiple expert networks can operate simultaneously without waiting for each other, this shortens the overall computation time and improves computational efficiency. This method is particularly suitable for processing large-scale molecular structure data in the embodiments of this application.

[0102] Step 130: Input the graph feature data and allocation results into at least one expert model to perform data prediction and obtain the predicted potential energy and predicted force field corresponding to each peptide unit.

[0103] In one embodiment, reference is made to Figure 1 After obtaining the allocation results, the graph feature data and allocation results are sent into memory through the PCIe interface. Next, each expert model selects the corresponding graph feature data according to the allocation results to perform data prediction, thereby obtaining the predicted potential energy and predicted force field corresponding to each peptide unit.

[0104] For peptide units, the total molecular energy includes potential energy, which reflects the energy of interactions within the peptide unit and with its surrounding environment. This potential energy includes van der Waals potential energy, chemical bond stretching potential energy, bond angle change potential energy, torsional potential energy, and Coulomb interaction potential energy. The force field is fundamental to molecular dynamics simulations. It describes the interactions between atoms and molecules within a molecule using a set of parameters and functions, transforming the interactions between atoms into a mathematical representation of forces and energy, which can be calculated from the potential energy.

[0105] In one embodiment, reference is made to Figure 8 , Figure 8 This application provides a flowchart of the process of inputting graph feature data and allocation results into at least one expert model for data prediction to obtain the predicted potential energy and predicted force field corresponding to the peptide unit. The flowchart specifically includes the following steps:

[0106] Step 810: Based on the allocation results, obtain at least one graphical feature data corresponding to each expert model as representation data.

[0107] In one embodiment, for each expert model, its corresponding graph feature data is used as representation data.

[0108] Step 820: Use an expert model to predict the characterization data, obtain the simulated trajectory position corresponding to the three-dimensional coordinates of the peptide unit, and obtain the predicted potential energy and predicted force field of the peptide unit based on the simulated trajectory position.

[0109] In one embodiment, the expert model can perform molecular dynamics simulations, that is, it can simulate the motion of peptide units at multiple time steps, during which corresponding energy changes occur. (Refer to...) Figure 9 , Figure 9 This application provides a flowchart of a process for predicting the simulated trajectory position of a peptide unit's three-dimensional coordinates using an expert model, which includes the following steps:

[0110] Step 910: Based on the initial velocity and mass information of the protein, obtain the velocity and acceleration data of each peptide unit at different time steps.

[0111] In one embodiment, a micro-canonical ensemble (NVE) is used for molecular dynamics simulations. This system's simulation scenario is set with a defined number of particles (N), volume (V), and energy (E) during protein dynamics simulations. The protein's dissolution system during the simulation is defined as sodium chloride solvent, with added ions to ensure electroneutrality. An initial temperature T0 is defined. Before the simulation, the initial energy of the system is minimized using gradient descent, thereby eliminating unfavorable interatomic contacts and collisions, bringing the system to a pre-equilibrium state. At the start of the simulation, the initial velocity and mass information of the protein are acquired or defined.

[0112] In one embodiment, after entering the dynamic simulation process, the simulation is performed according to a preset number of time steps, and the motion state of the peptide unit will change accordingly after each time step.

[0113] Therefore, at each time step, based on Newton's theorem and the initial velocity information of the protein, the velocity and acceleration data of each peptide unit at different time steps are obtained, expressed as:

[0114]

[0115] in, This represents the initial velocity information, and Δt represents the time change corresponding to the t-th time step. This represents the acceleration data at that time step. This represents the velocity data at that time step, where both acceleration and velocity data are vectors. Represents a vector field.

[0116] Step 920: Update the three-dimensional coordinates of the peptide unit at least for each time step using velocity and acceleration data.

[0117] In one embodiment, the three-dimensional coordinates of the peptide unit at each time step can be calculated using velocity and acceleration data. Taking the x-coordinate of the residue as an example, the updated three-dimensional coordinates of the peptide unit are expressed as follows:

[0118]

[0119] Among them, R bx R represents the initial position of the residue's x-coordinate. bx_new This indicates the updated x-coordinate position of the residue. It can be understood that the position updates for all three-dimensional coordinates within the peptide unit are performed using the above calculation method.

[0120] Step 930: After a preset number of time steps of dynamic simulation, obtain the three-dimensional coordinates of the peptide unit corresponding to the last time step as the simulation trajectory position.

[0121] In one embodiment, the preset number of time steps can be set according to actual needs, that is, the time required for molecular dynamics simulation. Following the aforementioned process, the three-dimensional position of each peptide unit is updated at each time step, and the three-dimensional coordinates of the peptide unit corresponding to the last time step are used as the simulation trajectory position. Alternatively, visualization software, such as PyMOL, can be used to display the position of the peptide unit at each time step, thereby visualizing the molecular dynamics simulation process.

[0122] Once the simulated trajectory position is obtained, the predicted potential energy and predicted force field corresponding to the peptide unit's movement to that position can be derived. Alternatively, the predicted force field can be obtained by differentiating the predicted potential energy with respect to the simulated trajectory position.

[0123] Step 830: Summarize the predicted potential energy and predicted force field outputs from all expert models.

[0124] In one embodiment, following the above process, each expert model predicts the predicted potential energy and predicted force field corresponding to the corresponding characterization feature. Therefore, by summing up the predicted potential energy and predicted force field of all expert models, the predicted potential energy and predicted force field of all peptide units can be obtained.

[0125] In this embodiment, the dynamic simulation prediction model undergoes heterogeneous inference. The graph network structure, primarily used for feature extraction, is deployed in GPU memory for efficient data reading and feature extraction. The hybrid expert model is deployed in CPU memory, utilizing multiple CPU cores for parallel inference. This fully leverages the processing advantages of different computing resources in the heterogeneous cluster, accelerating the inference speed of the hybrid expert model. Furthermore, because the hybrid expert model is deployed in memory, its scale can be increased, improving inference accuracy.

[0126] The training process of the dynamic simulation prediction model in the embodiments of this application is described below.

[0127] In one embodiment, reference is made to Figure 10 , Figure 10 This is a schematic diagram of the training process of the dynamic simulation prediction model provided in the embodiments of this application, including the following steps:

[0128] Step 1010: Construct the sample dataset.

[0129] In one embodiment, manually proofread protein data and a sample dataset can be downloaded from an open-source database. For example, the Swiss-Prot dataset contains 572,619 protein records, including protein type, amino acid chain, and atomic three-dimensional coordinates. Data filtering can also be performed to retain proteins with a length L < 500, resulting in a total of M = 460,587 protein sequences.

[0130] Next, in this embodiment of the application, multiple training samples are generated based on the acquired data to form a sample dataset. The training samples include training protein sequences and training labels. The segmentation method and data dimensions of the training protein sequences are consistent with those of the protein sequences to be tested. Unlike the protein sequences to be tested, the training protein sequences also include training labels, which include potential energy labels and force field labels.

[0131] The potential energy label and force field label can be obtained using numerical calculation methods based on density functional theory (DFT). The potential energy label is represented as V. b The force field label is represented as Therefore, the sample dataset Represented as:

[0132]

[0133] b∈[1,L-1]}

[0134] Where M represents the number of training samples in the sample dataset, P j Let P represent the j-th training sample, and b represent the training sample P.j The b-th peptide unit represents These are the three-dimensional coordinates of carbon atoms at different positions in the b-th peptide unit. These are the three-dimensional coordinates of the oxygen atom in the b-th peptide unit. These are the three-dimensional coordinates of the nitrogen atom in the b-th peptide unit. These are the three-dimensional coordinates of the residues connected to the b-th peptide unit. The three-dimensional coordinates are the x, y, and z coordinate components in three-dimensional space. The tag representing the potential energy of the b-th peptide unit. This represents the force field label of the b-th peptide unit.

[0135] Step 1020: Input the training samples into the dynamic simulation prediction model for data processing to obtain the training prediction potential energy and the training prediction force field.

[0136] In one embodiment, the training epochs of the dynamic simulation prediction model are defined as Epoch = 50, and in each epoch, the training epochs of the model are... All training samples are randomly shuffled, and the batch size BS = 16 is defined for each epoch of training. Assuming M = 460587, each epoch contains M / BS = 28786 iterations. The model parameter updates use the AdamW optimizer with an initial learning rate of 0.001. The learning rate decays through linear decay over 50 epochs, eventually reaching a minimum learning rate of 0.00001. Simulating the dynamics of Endonuclease Htp3 protein (a macromolecule) is used as an example. Htp3 protein contains 211 amino acids, or 210 peptide units. 290 empty nodes are filled to a total of 500 peptide units. The dynamics are simulated using Amber20 software. The training termination condition is defined as reaching the total simulation duration, set to 10 ns. Assuming a time step of 1 fs, the total number of iterations during the simulation is Iter = 10. 7 .

[0137] Since GPU memory is typically less than 80GB, it is difficult to deploy hybrid expert models (100GB level) on a single GPU device when using a single GPU for inference. Considering that the memory (TB level) of heterogeneous clusters is often idle, this application embodiment deploys the hybrid expert model in memory to realize the deployment and inference of large-scale dynamic simulation prediction models on a single heterogeneous machine. This can fully utilize heterogeneous computing power and the multi-core computing of the CPU to accelerate the simulation process through parallelization.

[0138] During training, the training samples are input into the dynamic simulation prediction model for data processing, and the training prediction potential energy is output. Then, the training prediction potential energy is differentiated with respect to the three-dimensional coordinates of the peptide unit to obtain the training prediction force field of each peptide unit and the atoms it contains.

[0139] Step 1030: Calculate the potential energy loss value based on the potential energy label and the training predicted potential energy, calculate the force field loss value based on the force field label and the training predicted force field, and calculate the total loss value based on the potential energy loss value and the force field loss value.

[0140] In one embodiment, the loss function needs to include potential energy loss values ​​related to potential energy. Force field loss value related to force field Then, calculate the total loss value based on the potential energy loss value and the force field loss value. The total loss is expressed as:

[0141] Step 1040: Adjust the model parameters of the dynamic simulation prediction model according to the total loss value until the training termination condition is met, and obtain the trained dynamic simulation prediction model.

[0142] In one embodiment, after all rounds of training are completed, a trained dynamic simulation prediction model is obtained. When the protein sequence to be tested is input, the model can predict the predicted potential energy and predicted force field of each peptide unit in the protein.

[0143] In one embodiment, to verify the inference performance of the dynamic simulation prediction model in this application embodiment, the time t of the inference process of a single dynamic simulation prediction model is recorded. mlff =12.8s, and the time t calculated using traditional DFT. dft =1.31h. It can be seen that, compared with the DFT method, the dynamic simulation prediction model of this application improves the inference speed by two orders of magnitude, or more than 300 times, in predicting potential energy and force field.

[0144] In one embodiment, the prediction loss (MAE) for potential energy in this application embodiment is calculated based on the potential energy label. mlff =0.132kcal mol -1 And the prediction loss MAE corresponding to the DFT method cm =2.978 kcal mol -1 As can be seen, the dynamic simulation prediction model of this application improves the simulation accuracy by an order of magnitude, that is, more than 20 times.

[0145] The technical solution provided in this application embodiment is executed by a pre-trained dynamic simulation prediction model. This model includes a graph network structure and a gating structure deployed on a GPU device, and at least one parallel expert model deployed on a CPU device. It acquires the protein sequence corresponding to the protein being tested, which includes multiple peptide units. The protein sequence is input into the graph network structure for feature extraction, obtaining graph feature data corresponding to each peptide unit. The graph feature data is then input into the gating structure for feature processing to obtain the allocation result corresponding to each expert network. The graph feature data and allocation results are then input into at least one expert model for data prediction, obtaining the predicted potential energy and predicted force field corresponding to each peptide unit. In this application embodiment, the dynamic simulation prediction model performs heterogeneous inference. The graph network structure, mainly used for feature extraction, is deployed in GPU memory for efficient data reading and feature extraction. The hybrid expert model is deployed in CPU memory, utilizing multiple CPU cores for parallel inference. This fully leverages the processing advantages of different computing resources in the heterogeneous cluster, accelerating the inference speed of the hybrid expert model. Furthermore, since the hybrid expert model is deployed in memory, its scale can be increased, improving inference accuracy.

[0146] This application also provides a molecular dynamics simulation and prediction device that can implement the above-described molecular dynamics simulation and prediction method, referring to... Figure 11 The device includes:

[0147] Data acquisition module 1110: used to acquire the protein sequence to be tested, which includes multiple peptide units.

[0148] Feature extraction module 1120: It is used to input the protein sequence to be tested into the graph network structure for feature extraction, obtain the graph feature data corresponding to each peptide unit, and input the graph feature data into the gating structure for feature processing to obtain the allocation result corresponding to each expert network.

[0149] Expert prediction module 1130: used to input graph feature data and allocation results into at least one expert model to perform data prediction, and obtain the predicted potential energy and predicted force field corresponding to each peptide unit.

[0150] The specific implementation of the molecular dynamics simulation and prediction device in this embodiment is basically the same as the specific implementation of the molecular dynamics simulation and prediction method described above, and will not be repeated here.

[0151] This application also provides an electronic device, including:

[0152] At least one memory;

[0153] At least one processor;

[0154] At least one program;

[0155] The program is stored in a memory, and the processor executes the at least one program to implement the molecular dynamics simulation and prediction method described above in this application. The electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0156] Please see Figure 12 , Figure 12 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0157] The processor 1201 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0158] The memory 1202 can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1202 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1202 and is called and executed by the processor 1201 using the molecular dynamics simulation prediction method of the embodiments of this application.

[0159] The input / output interface 1203 is used to implement information input and output;

[0160] Communication interface 1204 is used to enable communication and interaction between this device and other devices. Communication can be achieved via wired means (e.g., USB, Ethernet cable) or wireless means (e.g., mobile network, Wi-Fi, Bluetooth).

[0161] Bus 1205 transmits information between various components of the device (e.g., processor 1201, memory 1202, input / output interface 1203, and communication interface 1204);

[0162] The processor 1201, memory 1202, input / output interface 1203 and communication interface 1204 are connected to each other within the device via bus 1205.

[0163] This application embodiment also provides a storage medium that stores a computer program, which, when executed by a processor, implements the above-described molecular dynamics simulation and prediction method.

[0164] Memory, as a non-transitory storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0165] The molecular dynamics simulation prediction method, apparatus, device, and storage medium proposed in this application are executed by a pre-trained dynamics simulation prediction model. This model includes a graph network structure and a gating structure deployed on a GPU device, and at least one parallel expert model deployed on a CPU device. The method involves acquiring the protein sequence to be tested, which includes multiple peptide units. The protein sequence is input into the graph network structure for feature extraction, yielding graph feature data for each peptide unit. This graph feature data is then input into the gating structure for feature processing to obtain the allocation result for each expert network. Finally, the graph feature data and allocation results are input into at least one expert model for data prediction, yielding the predicted potential energy and predicted force field for each peptide unit. In this application, the dynamics simulation prediction model undergoes heterogeneous inference. The graph network structure, primarily used for feature extraction, is deployed in GPU memory for efficient data reading and feature extraction. The hybrid expert model is deployed in CPU memory, utilizing multiple CPU cores for parallel inference. This fully leverages the processing advantages of different computing resources in a heterogeneous cluster, accelerating the inference speed of the hybrid expert model. In addition, since hybrid expert models are deployed in memory, they can be scaled up and their inference accuracy improved.

[0166] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0167] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0168] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0169] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0170] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0171] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0172] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0173] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0174] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0175] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0176] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A molecular dynamics simulation prediction method, characterized in that, The method is executed by a pre-trained dynamics simulation prediction model, which includes a graph network structure and a gating structure deployed on a GPU device, and at least one parallel expert model deployed on a CPU device. The graph network structure includes a graph convolutional network and a linear transformation network. Obtain the sequence of the protein to be tested corresponding to the protein, wherein the sequence of the protein to be tested includes multiple peptide units; The graph convolutional network is used to perform graph convolution operations on each peptide unit in the protein sequence to be tested to obtain first feature data. The linear transformation network is then used to perform convolution operations on the first feature data to obtain multiple convolutional feature data. The multiple convolutional feature data are concatenated and then split into query data, key data, and value data. The query data and the key data are then processed with activation functions to obtain intermediate query data and intermediate key data. The intermediate query data, the intermediate key data, and the value data are then matrix multiplied and then convolutionally processed to obtain graph feature data corresponding to each peptide unit. The graph feature data is then input into the gating structure for feature processing to obtain the allocation result corresponding to each expert model. The graph feature data and the allocation result are input into the at least one expert model for data prediction to obtain the predicted potential energy and predicted force field corresponding to each peptide unit.

2. The molecular dynamics simulation and prediction method according to claim 1, characterized in that, The acquisition of the protein sequence to be tested includes: Obtain the initial protein data corresponding to the protein, and divide the initial protein data with peptide bonds as the center to obtain multiple peptide units; The three-dimensional coordinates of carbon atoms, oxygen atoms, nitrogen atoms, and residues in each peptide unit are obtained as the three-dimensional coordinates of the peptide unit, thus obtaining the protein sequence to be tested.

3. The molecular dynamics simulation and prediction method according to claim 2, characterized in that, The step of inputting the graph feature data and the allocation result into the at least one expert model for data prediction to obtain the predicted potential energy and predicted force field corresponding to the peptide unit includes: Based on the allocation result, at least one graph feature data corresponding to each expert model is obtained as representation data; The expert model is used to predict the characterization data to obtain the simulated trajectory position corresponding to the three-dimensional coordinates of the peptide unit, and the predicted potential energy and predicted force field of the peptide unit are obtained based on the simulated trajectory position. Summarize the predicted potential energy and predicted force field outputs from all the expert models.

4. The molecular dynamics simulation and prediction method according to claim 3, characterized in that, The step of using the expert model to predict the characterization data and obtain the simulated trajectory position corresponding to the three-dimensional coordinates of the peptide unit includes: Based on the initial velocity information of the protein, velocity and acceleration data of each peptide unit at different time steps are obtained; The three-dimensional coordinates of the peptide unit corresponding to each time step are updated at least using the velocity data and the acceleration data; After a preset number of time steps of dynamic simulation, the three-dimensional coordinates of the peptide unit corresponding to the last time step are obtained as the simulated trajectory position.

5. The molecular dynamics simulation and prediction method according to claim 1, characterized in that, The training process of the dynamic simulation prediction model includes the following steps: Construct a sample dataset, which includes multiple training samples, each training sample including training protein sequences and training labels, the training labels including potential energy labels and force field labels; The training samples are input into the dynamic simulation prediction model for data processing to obtain the training prediction potential energy and the training prediction force field. The potential energy loss value is calculated based on the potential energy label and the training predicted potential energy; the force field loss value is calculated based on the force field label and the training predicted force field; and the total loss value is calculated based on the potential energy loss value and the force field loss value. The model parameters of the dynamic simulation prediction model are adjusted according to the total loss value until the training termination condition is met, thus obtaining the trained dynamic simulation prediction model.

6. A molecular dynamics simulation and prediction device, characterized in that, The device is executed by a pre-trained dynamics simulation prediction model, which includes a graph network structure and a gating structure deployed on a GPU device, and at least one parallel expert model deployed on a CPU device. The graph network structure includes a graph convolutional network and a linear transformation network. Data acquisition module: used to acquire the sequence of the protein to be tested, which includes multiple peptide units; Feature extraction module: This module uses the graph convolutional network to perform graph convolution operations on each peptide unit in the protein sequence to be tested, obtaining first feature data; it then uses the linear transformation network to perform convolution operations on the first feature data, obtaining multiple convolutional feature data; it concatenates the multiple convolutional feature data and then splits them into query data, key data, and value data; it applies the query data and key data to activation functions to obtain intermediate query data and intermediate key data; it performs matrix multiplication on the intermediate query data, intermediate key data, and value data, and then performs convolution operations to obtain graph feature data corresponding to each peptide unit; finally, it inputs the graph feature data into the gating structure for feature processing to obtain the allocation result corresponding to each expert model. Expert prediction module: used to input the graph feature data and the allocation result into the at least one expert model to perform data prediction, and obtain the predicted potential energy and predicted force field corresponding to each peptide unit.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the molecular dynamics simulation and prediction method according to any one of claims 1 to 5.

8. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the molecular dynamics simulation and prediction method according to any one of claims 1 to 5.

Images