Trajectory prediction model construction method, trajectory prediction method, device, storage medium and computer program product

Abstract

This application discloses a trajectory prediction model construction method, trajectory prediction method, device, storage medium, and computer program product, relating to the fields of artificial intelligence and computational structural biology. The trajectory prediction model construction method includes: acquiring biomacromolecule trajectories and their physical timestamps; adding noise to the biomacromolecule trajectories based on a preset noise scale to generate noisy biomacromolecule trajectories; determining refined characterization of the biomacromolecules; predicting the biomacromolecules using an initial trajectory prediction model based on the noisy biomacromolecule trajectories, the preset noise scale, the physical timestamps, and the refined characterization to obtain predicted coordinates; and training the initial trajectory prediction model based on the predicted coordinates to obtain a target trajectory prediction model. By implementing this invention, the dynamic trajectories of biomacromolecules can be generated through machine learning.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and computational structural biology, and in particular to a trajectory prediction model construction method, trajectory prediction method, device, storage medium and computer program product. Background Technology

[0002] With the rapid development of protein structure prediction technology, research on the prediction of biomolecular dynamic trajectories has shifted from molecular dynamics simulations to machine learning approaches in recent years. However, existing machine learning models, when predicting the dynamic trajectories of biomolecular molecules, limit their input data only to the atomic structure of the biomolecular molecules. This results in a lack of temporal dimension information in existing machine learning models, which can only output static structures or conformational distributions and cannot characterize the continuous dynamic behavior of biomolecular evolution over time. Summary of the Invention

[0003] The technical problem to be solved by this application is to propose a trajectory prediction model construction method, trajectory prediction method, device, storage medium and computer program product to address the shortcomings of the existing technology.

[0004] To achieve the above objectives, this application proposes a trajectory prediction model construction method, which includes: Obtain the trajectory of a biomacromolecule, the physical timestamp of the biomacromolecule trajectory, the amino acid sequence of the biomacromolecule protein, and the ligand small molecule that binds to the protein; Noise is added to the biomacromolecule trajectory based on a preset noise scale to generate a noisy biomacromolecule trajectory. The refined characterization of the biomacromolecule was obtained based on the amino acid sequence and the ligand small molecule. Based on the noisy biomacromolecule trajectory, the preset noise scale, the physical timestamp of the biomacromolecule trajectory, and the refined characterization of the biomacromolecule, the predicted coordinates of the biomacromolecule are obtained by predicting using an initial trajectory prediction model, wherein the initial trajectory prediction model includes a diffusion encoder and a diffusion decoder. The initial trajectory prediction model is trained based on the predicted coordinates to obtain the target trajectory prediction model.

[0005] In one feasible implementation, the step of adding noise to the biomacromolecule trajectory based on a preset noise scale to generate a noisy biomacromolecule trajectory includes: Set a conditional frame for the biomacromolecule trajectory, and set the other trajectory frames in the biomacromolecule trajectory other than the conditional frame as the target frame; Based on the preset noise scale, noise is added to the conditional frame and each of the target frames to obtain a noisy conditional frame and each noisy target frame. Based on the noisy conditional frames and each of the noisy target frames, a noisy biomacromolecule trajectory is generated.

[0006] In one feasible implementation, the refining and characterization of the biomacromolecule based on the amino acid sequence and the ligand small molecule includes: Structural templates for each type of amino acid, structural templates for the ligand small molecules, and atomic type characteristics of the biomacromolecules are obtained from a pre-set database. The amino acid sequence is processed to obtain the sequence alignment of the amino acid sequence; The atomic topology of the ligand small molecule is obtained based on the ligand small molecule; The amino acid sequence, the atomic topology, the sequence alignment, the structural template of each type of amino acid, the structural template of the ligand small molecule, and the atomic type features of the biomacromolecule are passed through a preset feature encoder to obtain a single representation of each word. Based on the single representation of each lexical unit, a refined representation of the biological macromolecule is obtained.

[0007] In one feasible implementation, obtaining the refined characterization of the biomolecule based on the single characterization of each lexical unit includes: Pair all lexical units together to obtain several lexical pairs; After linear transformation, the pair representations of each lexical pair are obtained; Triangular updates and attention interactions are performed on the single representations of all lexical units and the pair representations of all lexical pairs to obtain the refined single representations and refined pair representations of the biomacromolecules. The refined characterization of the biomacromolecule is obtained based on the single characterization of the refined biomacromolecule and the pair characterization of the refined biomacromolecule.

[0008] In one feasible implementation, the step of predicting the predicted coordinates of the biomacromolecule based on the noisy biomacromolecule trajectory, the preset noise scale, the physical timestamp of the biomacromolecule trajectory, and the refined characterization of the biomacromolecule through an initial trajectory prediction model includes: The noisy biomacromolecule trajectory, the preset noise scale, the noisy biomacromolecule trajectory, the refined characterization of the biomacromolecule, and the atomic type features of the biomacromolecule are passed through the diffusion encoder to obtain the latent representation of each trajectory frame. One trajectory frame contains multiple atoms, and the latent representation of one trajectory frame is composed of the latent vectors of multiple atoms. For each atom in each trajectory frame, the spatial attention result for each atom is obtained through the spatial attention mechanism of the diffusion encoder based on the latent vector of each atom; For each identical atom in any two trajectory frames, the temporal attention result for each identical atom is obtained through the temporal attention mechanism of the diffusion encoder based on the latent vector of each identical atom and the physical timestamp of the two trajectory frames. The spatial attention result of each atom and the temporal attention result of the same atom are fused to obtain the fused potential representation of each trajectory frame; The fused latent representation of each trajectory frame is decoded by the diffusion decoder to obtain the predicted coordinates of each trajectory frame.

[0009] In one feasible implementation, obtaining the spatial attention result for each atom in each trajectory frame, based on the latent vector of each atom through the spatial attention mechanism of the diffusion encoder, includes: A linear transformation is performed on the potential vector of each first atom in each trajectory frame to obtain the query vector of each trajectory frame, wherein the first atom is any atom in each trajectory frame; A linear transformation is performed on the potential vector of the second atom in each trajectory frame to obtain the bond vector and value vector of each trajectory frame, wherein the second atom is any atom other than the first atom in each trajectory frame; A linear transformation was performed on the refined characterization of the biomacromolecules to obtain the spatial bias term; The spatial attention coefficient of each atom is calculated based on the spatial bias term, the query vector of each trajectory frame, the first atom of each trajectory frame, the second atom of each trajectory frame, and the third atom associated with the first atom. The spatial attention result for each atom is obtained based on the spatial attention coefficient of each atom and the value vector of each trajectory frame.

[0010] In one feasible implementation, for each identical atom in any two trajectory frames, the temporal attention result for each identical atom is obtained through the temporal attention mechanism of the diffusion encoder based on the latent vector of each identical atom and the physical timestamps of the two trajectory frames, including: From the noisy biomacromolecule trajectories, select all first trajectory frames and second trajectory frames that have the same atom, wherein the number of the same atom is at least one; A linear transformation is performed on the potential vector of each identical atom in each of the first trajectory frames to obtain the query vector of each identical atom in the first trajectory frame; A linear transformation is performed on the latent vector of each identical atom in each of the second trajectory frames to obtain the bond vector and value vector of each identical atom in the second trajectory frame; A time bias term is generated based on the time attention mechanism, the physical timestamp of the first trajectory frame, and the physical timestamp of the second trajectory frame; Calculate the time attention coefficient for each identical atom based on the time bias term and each of the query vectors; The time attention result for each of the same atoms is obtained based on the time attention coefficient of each atom and each value vector.

[0011] In one feasible implementation, training the initial trajectory prediction model based on the predicted coordinates to obtain the target trajectory prediction model includes: Obtain the true coordinates of the biomolecule's trajectory; The prediction error between the predicted coordinates and the actual coordinates of the biomolecule is calculated using a preset loss function. The loss function is optimized using a preset diffusion model technique based on the prediction error, and the target trajectory prediction model is obtained when the loss function converges.

[0012] Furthermore, to achieve the above objectives, this application also proposes a trajectory prediction method, which is implemented through a target trajectory prediction model constructed by a trajectory prediction model construction method. The trajectory prediction method includes: The initial trajectory and initial trajectory information of a known biomacromolecule are preset, wherein the initial trajectory includes a preset condition frame and a set of target frames to be generated, and the initial trajectory information includes the coordinates of the condition frame, the number of target frames to be generated, and the physical timestamp of the initial trajectory; Obtain the amino acid sequence of the known biological macromolecule protein and the ligand small molecule that binds to the protein; The coordinates of the target frame set to be generated are initialized according to a preset noise scale to obtain the initial coordinates of the target frame set to be generated; The refined characterization of the known biomacromolecules was obtained based on the amino acid sequence and the ligand small molecule. Based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, the target trajectory prediction model iteratively denoises the initial coordinates of the target frame set to be generated, thereby generating the predicted coordinates of the target frame set to be generated. Based on the coordinates of the conditional frame and the predicted coordinates of each set of target frames to be generated, the target trajectory of the known biomacromolecule is generated.

[0013] In one feasible implementation, the step of iteratively denoising the initial coordinates of the target frame set to be generated using the target trajectory prediction model based on the initial coordinates of the target frame set to be generated, the coordinates of the conditional frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, to generate the predicted coordinates of the target frame set to be generated, includes: Based on the initial coordinates of the target frame set to be generated, the coordinates of the conditional frame, the first preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, the first coordinate estimate of the target frame set to be generated is obtained through the target trajectory prediction model. The first coordinate estimate of each set of target frames to be generated, the initial coordinates of the target frames to be generated, and the second preset noise scale are used to generate the first predicted coordinates of the set of target frames to be generated through a preset diffusion model. The process involves repeatedly updating the initial coordinates of the target frame set to be generated based on the first predicted coordinates, and updating the first preset noise scale based on the second preset noise scale. Then, using the target trajectory prediction model, the initial coordinates of the target frame set to be generated, the coordinates of the conditional frame, the first preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules are used to obtain the first coordinate estimate of the target frame set to be generated. The second preset noise scale is then updated based on the third preset noise scale. Finally, an iterative denoising step is performed, using the first coordinate estimates of each target frame to be generated, the initial coordinates of the target frame to be generated, and the second preset noise scale, through a preset diffusion model to generate the first predicted coordinates of the target frame to be generated. This process continues until the number of iterations of the iterative denoising step reaches a preset iteration stop condition, thus obtaining the predicted coordinates of the target frame set to be generated.

[0014] In one feasible implementation, the initial coordinates of the target frame set to be generated are iteratively denoised using the target trajectory prediction model based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecule, to generate the predicted coordinates of the target frame set to be generated; based on the coordinates of the condition frame and the predicted coordinates of the target frame set to be generated, the target trajectory of the known biomacromolecule is generated, including: If the difference in physical timestamps of the target frame set to be generated is greater than a preset threshold, the physical timestamps of the target frames to be generated are fragmented to obtain multiple fragmented target frames and multiple fragmented physical timestamps of the fragmented target frames. For each segmented target frame, based on the segmented target frame and its segmented physical timestamp, the sparse predicted coordinates of each segmented target frame are obtained through the target trajectory prediction model. For each intermediate target frame between adjacent fragmented target frames, a target condition frame and an intermediate target frame to be generated are set in each intermediate target frame, and a physical timestamp is set for the intermediate target frame. Each adjacent fragmented target frame is used as the target condition frame, the sparse prediction coordinates corresponding to the adjacent fragmented target frames are used as the coordinates of the target condition frame, and the other target frames in each intermediate target frame except the target condition frame are set as intermediate target frames to be generated. Based on each target condition frame, the coordinates of each target condition frame, each intermediate target frame to be generated, and the physical timestamp of the intermediate target frame, the predicted coordinates of each intermediate target frame to be generated are generated by the target trajectory prediction model. Based on the coordinates of each target condition frame and the predicted coordinates of each intermediate target frame, the coordinates of each intermediate target frame to be generated are generated. Based on the coordinates of all intermediate frames to be generated, the target trajectory of the known biomacromolecule is obtained.

[0015] In one feasible implementation, for each segmented target frame, the step of obtaining sparse predicted coordinates for each segmented target frame based on the segmented target frame and its segmented physical timestamp through the target trajectory prediction model includes: For any segmented target frame, the segmented physical timestamp of the segmented target frame is used as the physical timestamp of the initial trajectory, the segmented target frame is used as the initial trajectory, and the process returns to execute the step of generating the predicted coordinates of the target frame set to be generated through the target trajectory prediction model based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, until the predicted coordinates of all segmented target frames are generated; The sparse predicted coordinates of each of the target fragments are obtained based on the predicted coordinates of all the target fragments.

[0016] In one feasible implementation, the step of generating predicted coordinates for each intermediate target frame to be generated using the target trajectory prediction model based on each target condition frame, the coordinates of each target condition frame, each intermediate target frame to be generated, and the physical timestamp of the intermediate target frame includes: For any intermediate target frame, the coordinates of the intermediate target frame to be generated are initialized based on the preset noise scale to obtain the initial coordinates of each intermediate target frame to be generated; Based on the initial coordinates of the intermediate target frame to be generated, the coordinates of the target condition frame, the preset noise scale, the physical timestamp of the intermediate target frame to be generated, and the refined characterization of the known biomacromolecules, the predicted coordinates of each remaining intermediate target frame to be generated are generated using the target trajectory prediction model. In addition, to achieve the above objectives, this application also proposes a trajectory prediction model construction device, which includes: a training data preprocessing module, a prediction coordinate generation module, and a model training module; The training data preprocessing module is used to acquire biomacromolecule trajectories, physical timestamps of the biomacromolecule trajectories, amino acid sequences of the biomacromolecule proteins, and small ligand molecules bound to the proteins; add noise to the biomacromolecule trajectories based on a preset noise scale to generate noisy biomacromolecule trajectories; and obtain refined characterization of the biomacromolecules based on the amino acid sequences and the small ligand molecules. The predicted coordinate generation module is used to predict the predicted coordinates of the biomacromolecule based on the noisy biomacromolecule trajectory, the preset noise scale, the physical timestamp of the biomacromolecule trajectory, and the refined characterization of the biomacromolecule through an initial trajectory prediction model, wherein the initial trajectory prediction model includes a diffusion encoder and a diffusion decoder. The model training module is used to train the initial trajectory prediction model based on the predicted coordinates to obtain the target trajectory prediction model.

[0017] In addition, to achieve the above objectives, this application also proposes a trajectory prediction device, which includes: a data preprocessing module, a coordinate prediction module, and a trajectory prediction module; The data preprocessing module is used to pre-set the initial trajectory and initial trajectory information of a known biomacromolecule, wherein the initial trajectory includes a pre-set condition frame and a set of target frames to be generated, and the initial trajectory information includes the coordinates of the condition frame, the number of target frames to be generated, and the set of physical timestamps of the initial trajectory; acquire the amino acid sequence of the known biomacromolecule protein and the ligand small molecule that binds to the protein; initialize the coordinates of the set of target frames to be generated according to a preset noise scale to obtain the initial coordinates of the set of target frames to be generated; and obtain the refined characterization of the known biomacromolecule based on the amino acid sequence and the ligand small molecule. The coordinate prediction module is used to iteratively denoise the initial coordinates of the target frame set to be generated based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, through the target trajectory prediction model, to generate the predicted coordinates of the target frame set to be generated. The trajectory prediction module is used to generate the target trajectory of the known biomacromolecule based on the coordinates of the condition frame and the predicted coordinates of the set of target frames to be generated.

[0018] Furthermore, to achieve the above objectives, this application also proposes a trajectory prediction model construction device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the trajectory prediction model construction method described above. This application also proposes a biomacromolecule trajectory prediction device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the trajectory prediction method described above.

[0019] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of the trajectory prediction model construction method described above, and / or when the computer program is executed by a processor, it implements the steps of the trajectory prediction method described above.

[0020] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the trajectory prediction model construction method as described above, and / or the computer program, when executed by a processor, implements the steps of the trajectory prediction method as described above.

[0021] One or more technical solutions proposed in this application have at least the following technical effects: This application provides a trajectory prediction model construction method, trajectory prediction method, device, storage medium, and computer program product. The trajectory prediction model construction method predicts the trajectory of the biomolecule based on the noisy biomolecule trajectory, the preset noise scale, the physical timestamp of the biomolecule trajectory, and the refined characterization of the biomolecule through an initial trajectory prediction model to obtain the predicted coordinates of the biomolecule; and trains the initial trajectory prediction model based on the predicted coordinates to obtain a target trajectory prediction model.

[0022] By incorporating physical timestamps of biomacromolecule trajectories into the training and generation of the target trajectory prediction model, the model learns temporal dimension information. Simultaneously, by introducing noisy biomacromolecule trajectories, a preset noise scale, and refined biomacromolecule representations, the model learns atomic structure dimension information, enabling the final target trajectory prediction model to directly characterize the continuous dynamic behavior of biomacromolecules evolving over time. This addresses the problem that existing machine learning models, which only use the atomic structure of biomacromolecules as input and lack temporal dimension information, can only output static structures or conformational distributions, making it difficult to characterize the continuous dynamic behavior of biomacromolecules evolving over time. Implementing this application provides analytical support for practical applications requiring high-throughput conformational sampling and drug molecule design based on biomacromolecule trajectories. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating the first embodiment of the trajectory prediction model construction method of this application; Figure 2This is a flowchart illustrating the second embodiment of the trajectory prediction method of this application. Figure 3 This is a flowchart illustrating the third embodiment of the trajectory prediction method of this application; Figure 4 This is a schematic diagram of the module structure of the trajectory prediction model construction device according to an embodiment of this application; Figure 5 This is a schematic diagram of the module structure of the trajectory prediction device according to an embodiment of this application; Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the trajectory prediction model construction method in the embodiments of this application; Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the trajectory prediction method in the embodiments of this application.

[0026] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0028] To better understand the technical solution of this application, it is necessary to clarify in advance that in the existing field of computational structural biology, the biological functions of biomolecules do not originate from unchanging static structures, but rather from their continuous dynamic transitions between different conformations. Key biological processes such as ligand binding and dissociation, signal transduction, and enzyme catalysis are all dominated by the kinetic rates of these conformational transitions. In recent years, structural prediction models, represented by AlphaFold, RoseTTAFold, and its open-source replica Protenix, have been able to accurately predict the three-dimensional ground-state structure of proteins and their complexes from amino acid sequences, greatly advancing the drug discovery process. However, the existing structural prediction models can only output static structures or conformational distributions and cannot correlate them in the time dimension. Consequently, they cannot generate time-correlated biomolecular dynamic trajectories based on the output static structures or conformational distributions using machine learning models.

[0029] Due to the lack of temporal correlation in the construction of biomolecular dynamic trajectories by the aforementioned machine learning models, the mainstream computational method for predicting biomolecular dynamic trajectories in the current field of computational structural biology is to use molecular dynamics (MD) simulation. However, molecular dynamics simulation requires femtosecond-level time steps to resolve high-frequency atomic vibrations, and simulating biologically relevant microsecond to millisecond timescales often takes several days or even months. This MD simulation method has excessive requirements for computational resources and time, and its portability is poor.

[0030] To address the problems existing in the prior art, this application proposes a trajectory prediction model construction method that can realize a time dimension, support arbitrary time resolution, and stably generate long-term biomolecular dynamic trajectories. The target trajectory prediction model can make up for the timeliness problem of existing MD simulations. At the same time, by training the trajectory prediction model, the problem that the static structure or conformational distribution output by the existing structure prediction model cannot generate biomolecular dynamic trajectories with temporal correlation can be solved.

[0031] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0032] In this embodiment, for ease of description, the following description uses a computer device as the execution subject. It should be noted that the execution subject in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or computer device capable of performing the above functions. The following description uses a computer device as an example to illustrate this embodiment and the subsequent embodiments.

[0033] The main solution of the first embodiment of this application is to propose a trajectory prediction model construction method, which trains an initial trajectory prediction model based on a diffusion model by constructing training data with time dimension and biomacromolecule atomic structure dimension, so as to train a target trajectory prediction model that can generate biomacromolecule dynamic trajectories with both time dimension and biomacromolecule atomic structure dimension information.

[0034] This application provides a method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the trajectory prediction model construction method of this application.

[0035] In this embodiment, the trajectory prediction model construction method includes steps S10 to S50: Step S10: Obtain the biomacromolecule trajectory, the physical timestamp of the biomacromolecule trajectory, the amino acid sequence of the biomacromolecule protein, and the ligand small molecule bound to the protein; Specifically, a segment of a known biomacromolecule trajectory containing several frames is obtained from a publicly available MD database, and the physical timestamp of the obtained biomacromolecule trajectory is determined, wherein each frame in the biomacromolecule trajectory corresponds to a physical timestamp.

[0036] When the acquired biomacromolecule trajectory contains T trajectory frames, the biomacromolecule trajectory is represented as follows: ,in This represents the all-atom three-dimensional coordinates of the i-th trajectory frame in the trajectory of a biomacromolecule. N is the number of atoms in a known biological macromolecule. The physical timestamp of this trajectory is represented as... ,in This represents the physical timestamp of the i-th frame in the trajectory of a biological macromolecule.

[0037] Known biomolecules obtained from publicly available MD databases include proteins and small ligand molecules that bind to those proteins. Therefore, when acquiring a biomolecule trajectory consisting of several frames, the amino acid sequence of the protein in the biomolecule trajectory and the small ligand molecules that bind to the protein can be acquired simultaneously.

[0038] Step S20: Add noise to the biomacromolecule trajectory based on a preset noise scale to generate a noisy biomacromolecule trajectory; Specifically, each trajectory frame of the biomacromolecule trajectory obtained in step S10 is denoised using a preset noise scale to generate a noisy biomacromolecule trajectory, wherein each trajectory frame of the noisy biomacromolecule trajectory contains noise.

[0039] In a feasible embodiment, step S20 may include steps A11 to A13: Step A11: Set a conditional frame for the biomacromolecule trajectory, and set the other trajectory frames in the biomacromolecule trajectory other than the conditional frame as the target frame.

[0040] Specifically, before adding noise to the biomolecular trajectory, at least one frame of the biomolecular trajectory must be set as a conditional frame, and the remaining trajectory frames of the biomolecular trajectory other than the set conditional frame must be set as target frames.

[0041] It should be noted that the purpose of setting condition frames and target frames is to enable the initial trajectory prediction model to be trained in the future, with the condition frames as the benchmark and the target frames as the standard for measuring the quality of the target trajectory prediction model. This allows the target trajectory prediction model to predict each target frame based on the condition frames, and the predicted target frames are temporally continuous under the constraint of physical timestamps, thereby achieving the goal of obtaining the predicted trajectory based on the prediction of each target frame.

[0042] In one specific embodiment, the first frame or the first and last two frames of the biomacromolecule trajectory can be set as conditional frames: when the first frame is used as a conditional frame, predicting all subsequent target frames through the first frame can enable the initial trajectory prediction model to learn the ability of "predicting the future with known past" (extrapolation / prediction), thereby achieving trajectory prediction; when the first and last frames of the biomacromolecule trajectory are set as conditional frames, predicting the target frames between the two through the first and last frames can enable the initial trajectory prediction model to learn the ability of "filling in the middle with known ends" (interpolation), thereby achieving trajectory interpolation prediction.

[0043] Additionally, it should be noted that the initial trajectory prediction model mentioned in this step and subsequent steps of this application is based on the diffusion model. Specifically, the initial trajectory prediction model can be based on the Protenix model, which can significantly reduce the computational cost of training during the training process of the initial trajectory prediction model.

[0044] Step A12: Noise is added to the conditional frame and each of the target frames based on the preset noise scale to obtain a noisy conditional frame and each noisy target frame; Specifically, when adding noise to any one of the conditional frames and target frames, noise is added to the conditional frames and target frames set in step A11 based on a preset noise scale to generate noisy conditional frames and noisy target frames.

[0045] For example, when adding noise to a conditional frame, in order to make it a clean visible frame (i.e. a clean, noise-free frame), the preset noise scale σ of the conditional frame can be set to σ=0 to obtain a noisy conditional frame.

[0046] Preferably, when adding noise to each target frame, a forward noise addition method is used. This involves randomly sampling noise from a standard Gaussian distribution, and then adding noise to each target frame based on the randomly sampled noise and a preset noise scale to obtain each noisy target frame. Specifically, for the atomic coordinate x of each target frame, noise ε is randomly sampled from a standard Gaussian distribution (N(0,I)) and combined with a preset noise scale. Noise is added to each target frame to obtain the full atomic coordinates of the noisy target frame. ,in , ε obeys N(0,I).

[0047] It should be noted that the preset noise scale σ follows a log-normal distribution, and ln(σ) follows an N(μ) distribution. noise , σ noise ), where μ noise The mean of the log-normal distribution is the value after taking the logarithm, and the average size of the preset noise scale σ is controlled. noise To control the discreteness and fluctuation range of the preset noise scale σ, the standard deviation of the log-normal distribution is calculated. In this application, μ... noise The optimal value for σ is -1.2, which results in a generally smaller generated noise scale. noise The preferred value is 1.5; when the preset noise scale σ is close to 0, the noisy trajectory frame is approximately a clean visible frame; when the preset noise scale σ is the maximum noise scale, the noisy trajectory frame is close to pure noise and can be regarded as a masked frame to be predicted.

[0048] Step A13: Generate noisy biomacromolecule trajectories based on the noisy conditional frames and each of the noisy target frames.

[0049] Specifically, by performing a time-series connection between each noisy conditional frame and each noisy target frame based on the physical timestamps corresponding to the noisy conditional frames and each noisy target frame, the trajectory of the noisy biomacromolecule can be obtained.

[0050] In this feasible implementation, by adding noise to each frame of the biomacromolecule trajectory, a noisy biomacromolecule trajectory is obtained, which can simulate the random perturbation in the process of protein molecule movement, expand the training sample space, enable the initial trajectory prediction model to learn the ability to restore the original coordinates of the trajectory frame under different noise intensities, and improve the rationality and generalization performance of the trajectory prediction model's output predicted coordinates.

[0051] Step S30: Refined characterization of the biomacromolecule is obtained based on the amino acid sequence and the ligand small molecule; Specifically, the refining and characterization of biological macromolecules are determined based on the amino acid sequence of proteins and the small ligands that bind to proteins.

[0052] It should be noted that the refined characterization of biological macromolecules is an important source of information on the atomic structure dimension of biological macromolecules. Refined characterization is used to fully explore the correlation features between amino acids and ligand atoms, integrating geometric and evolutionary priors.

[0053] In one feasible implementation, step S30 may include steps B11 to B15: Step B11: Obtain the structural templates of each type of amino acid, the structural templates of the ligand small molecules, and the atomic type characteristics of the biomacromolecules from a preset database; Specifically, structural templates for each type of amino acid, structural templates for ligands that bind to proteins, and atomic type characteristics of biological macromolecules are obtained from a pre-defined structural template database such as CCD (Chemical Component Dictionary).

[0054] Step B12: Process the amino acid sequence to obtain the sequence alignment of the amino acid sequence; Specifically, the amino acid sequence is obtained by using a multiple sequence alignment tool to perform multiple sequence alignment (MSA).

[0055] Step B13: Obtain the atomic topology of the ligand small molecule based on the ligand small molecule; Specifically, the small ligand molecules that bind to proteins are input into computational chemistry tools (such as RDkit) to obtain the atomic topology of the ligand molecules.

[0056] Step B14: The amino acid sequence, the atomic topology, the sequence alignment, the structural template of each type of amino acid, the structural template of the ligand small molecule, and the atomic type features of the biomacromolecule are passed through a preset feature encoder to obtain a single representation of each word. Specifically, the acquired amino acid sequences, atomic topologies of ligand small molecules, multiple sequence alignments of amino acid sequences, structural templates for each type of amino acid, structural templates for ligand small molecules, and atomic type features of the biomolecules are input into a pre-defined feature encoder to obtain a single representation for each token. Each token represents an atom of either an amino acid or a ligand small molecule. The pre-defined feature encoder can be implemented using the feature encoding module of an existing pre-trained Protenix model.

[0057] Step B15: Based on the single representation of each lexical unit, a refined representation of the biomolecule is obtained.

[0058] Specifically, the individual representations of each word are processed to obtain refined representations of biological macromolecules.

[0059] In one feasible implementation, step B15 may include steps C11-C14: Step C11: Pair all lexical units together to obtain several lexical pairs; Specifically, pair all tokens into any two pairs to obtain token pairs.

[0060] Step C12: After linear transformation, each of the lexical pairs is obtained as a pair representation of each lexical pair; Specifically, for each token pair, the single representations of the two tokens contained in the token pair are linearly transformed and then added together to obtain the pair representation of the token pair.

[0061] Step C13: Perform triangular updates and attention interactions on the single representations of all the lexical units and the pair representations of all the lexical pairs to obtain the refined single representations and refined pair representations of the biomacromolecules. Specifically, the Pairformer module of the existing Protenix model performs triangle updates and attention interactions on the single representations of all tokens and the paired representations of all token pairs, outputting refined single representations and refined paired representations containing rich geometry and evolutionary priors.

[0062] It should be noted that triangular update is used to propagate pairwise association information between tokens and optimize the pairwise representation; attention interaction realizes the aggregation of global information between the individual representations of each token and updates the individual representations. The refined representation obtained after the processing of triangular update and attention interaction integrates the features of the individual tokens themselves, the pairwise association features between tokens, and superimposed evolutionary priors and geometric constraints, possessing rich local atomic structure information and global atomic structure information.

[0063] Step C14: Based on the refined single characterization and the refined pair characterization of the biomacromolecule, the refined characterization of the biomacromolecule is obtained.

[0064] Specifically, refined single characterization and refined pair characterization are used together as refined characterization of biological macromolecules.

[0065] Step S40: Based on the noisy biomacromolecule trajectory, the preset noise scale, the physical timestamp of the biomacromolecule trajectory, and the refined characterization of the biomacromolecule, the predicted coordinates of the biomacromolecule are obtained by using an initial trajectory prediction model, wherein the initial trajectory prediction model includes a diffusion encoder and a diffusion decoder. Specifically, both the diffusion encoder and diffusion decoder of the initial trajectory prediction model include spatial attention and temporal attention mechanisms. The diffusion decoder also includes a decoding layer, which can be implemented using the alphafold3 architecture. The core of training the initial trajectory prediction model is to process the data input to the model. Temporal attention is used to read the context in the conditional frames and to perform denoising prediction on each noisy target frame in the noisy biomolecule trajectory to obtain the predicted coordinates of the biomolecule.

[0066] In this step, the noisy biomacromolecule trajectory, the preset noise scale, the physical timestamp of the biomacromolecule trajectory, and the refined characterization of the biomacromolecule are input into the initial trajectory prediction model, so that the initial trajectory prediction model outputs the denoised all-atom coordinates of each target frame. The output denoised all-atom coordinates are the predicted coordinates of the biomacromolecule.

[0067] In one feasible embodiment, step S40 may include steps D11-D15: Step D11: The noisy biomacromolecule trajectory, the preset noise scale, the noisy biomacromolecule trajectory, and the atom type features of the biomacromolecule are passed through the diffusion encoder to obtain the latent representation of each trajectory frame. One trajectory frame contains multiple atoms, and the latent representation of one trajectory frame is composed of the latent vectors of multiple atoms. Specifically, for each trajectory frame in the noisy biomacromolecule trajectory, the full atomic coordinates, noise scale, atomic type features, and refined characterization of the biomacromolecule of the trajectory frame are input into the diffusion encoder of the initial trajectory prediction model to obtain the atomic-level latent representation of each trajectory frame.

[0068] For example: For the first The trajectory frame will be the first one. The full atomic coordinates, noise scale, atomic type features, and refined characterization of biomacromolecules of the first trajectory frame are input into the diffusion encoder of the initial trajectory prediction model to obtain the first... The latent representation of a trajectory frame : ; Wherein, N is the first element in this biomacromolecule. The total number of atoms contained in each trajectory frame. A potential vector containing N atoms Indicates the first The latent vector of the k-th atom in each trajectory frame.

[0069] It should be noted that the latent vector contains the atom's local coordinates, atom / residue identity, ligand affiliation, spatial proximity, and conditional information from single / pair characterization.

[0070] Step D12: For each atom of each trajectory frame, obtain the spatial attention result of each atom based on the latent vector of each atom through the spatial attention mechanism of the diffusion encoder; Specifically, the latent vector of each atom is passed through the spatial attention mechanism of the diffusion encoder to obtain the spatial attention result of each atom.

[0071] For example, for the i-th trajectory frame, the latent vector of its k-th atom is... As the query input for the spatial attention mechanism, the potential representations of other atoms within the trajectory frame are used. As the inputs to the keys and values ​​in the spatial attention mechanism, the attention weights of this atom to other atoms in the same frame are calculated through the spatial attention mechanism, and the output result of this atom through the spatial attention mechanism is obtained. The calculation formula is as follows:

[0072] in, This represents the output of the spatial attention mechanism for the k-th atom in the i-th trajectory frame. Let be the spatial attention coefficient between the k-th atom and the l-th atom in the i-th trajectory frame; Let be the value vector of the l-th atom in the i-th trajectory frame.

[0073] The spatial attention coefficient between the k-th atom and the l-th atom in the i-th trajectory frame is calculated as follows:

[0074] in, and These are spatial bias terms, all obtained by a linear transformation of the characterization of biomacromolecules through refining. The dimensions of the query vector and key vector; Attention score based on content similarity This is the matrix transpose symbol; This indicates summation, which is performed on other atoms associated with the k-th atom. These other atoms can be all atoms within a preset distance range from the k-th atom in the i-th frame, atoms within the same amino acid as the k-th atom, and atoms chemically bonded to the k-th atom.

[0075] In one feasible implementation, step D12 may include steps E11 to E15: Step E11: Perform a linear transformation on the potential vector of each first atom in each trajectory frame to obtain the query vector of each trajectory frame, wherein the first atom is any atom in each trajectory frame; Specifically, any atom k in the i-th trajectory frame is selected as the first atom mentioned above, and the potential vector of the k-th atom in the i-th trajectory frame is... Perform a linear transformation to obtain the query vector of the k-th atom in the i-th trajectory frame. .in, The query is for the learnable linear transformation matrix.

[0076] Step E12: Perform a linear transformation on the potential vector of the second atom in each trajectory frame to obtain the bond vector and value vector of each trajectory frame, wherein the second atom is any atom other than the first atom in each trajectory frame; Specifically, any atom l in the i-th trajectory frame, excluding atom k, is selected as the second atom mentioned above, and the potential representation of the l-th atom in the i-th trajectory frame is obtained. Perform a linear transformation to obtain the bond vector of the l-th atom in the i-th trajectory frame. and the value vector of the l-th atom in the i-th trajectory frame .in, Let be the learnable linear transformation matrix of the key; This is a learnable linear transformation matrix with values ​​of . It should be noted that for each trajectory frame... , and All dimensions are the same.

[0077] Step E13: Perform a linear transformation on the refined characterization of the biomacromolecules to obtain the spatial bias term; Step E14: Calculate the spatial attention coefficient of each atom based on the spatial bias term, the query vector of each trajectory frame, the first atom of each trajectory frame, the second atom of each trajectory frame, and the third atom associated with the first atom. Specifically, the third atom mentioned above is the atom associated with the k-th atom in step D12. Based on the calculations obtained from E11-E14 above, the spatial attention coefficient of each atom is calculated using the spatial attention coefficient calculation formula given in formula (2) in step D12.

[0078] Step E15: Based on the spatial attention coefficient of each atom and the value vector of each trajectory frame, obtain the spatial attention result of each atom.

[0079] Specifically, based on the spatial attention coefficient of each atom and the value vector of each trajectory frame, the spatial attention result of each atom is calculated by formula (1) in step D12.

[0080] Step D13: For each identical atom in any two trajectory frames, based on the latent vector of each identical atom and the physical timestamp of the two trajectory frames, the temporal attention result of each identical atom is obtained through the temporal attention mechanism of the diffusion encoder. In one feasible implementation, step D13 may include steps F11 to F16: Step F11: Select all first trajectory frames and second trajectory frames with the same atom from the noisy biomacromolecule trajectories, wherein the number of the same atom is at least one. Specifically, from the noisy biomacromolecule trajectories, two trajectory frames with the same atom are selected and designated as the first trajectory frame and the second trajectory frame, respectively. The selected two trajectory frames must have at least one identical atom.

[0081] Step F12: Perform a linear transformation on the potential vector of each identical atom in each of the first trajectory frames to obtain the query vector of each identical atom in the first trajectory frame; Specifically, the first trajectory frame is represented by the i-th trajectory frame, and each identical atom in the first trajectory frame is represented by the k-th atom. Then, the potential vector of the k-th atom in the i-th trajectory frame is... Perform a linear transformation to obtain the query vector for each identical k-th atom in the i-th trajectory frame. ,in The query is for the learnable linear transformation matrix.

[0082] Step F13: Perform a linear transformation on the potential vector of each identical atom in the second trajectory frame to obtain the bond vector and value vector of each identical atom in the second trajectory frame; Specifically, if a second trajectory frame that has the same atoms as the first trajectory frame (the i-th trajectory frame) is represented as the j-th trajectory frame, then the potential representation of the k-th atom in the j-th trajectory frame is... Perform a linear transformation to obtain the bond vector of the k-th atom in the j-th trajectory frame. and the value vector of the kth atom in the j-th trajectory frame ,in and These are the learnable linear transformation matrices for the keys and values, respectively. It should be noted that for each trajectory frame... , and All dimensions are the same.

[0083] Step F14: Generate a time bias term based on the time attention mechanism, the physical timestamp of the first trajectory frame, and the physical timestamp of the second trajectory frame; Specifically, a time bias term is generated based on the learnable decay factor in the time attention mechanism, the physical timestamp of the first trajectory frame, and the physical timestamp of the second trajectory frame. The calculation formula for the time bias term is as follows:

[0084] in, λ represents the temporal bias terms corresponding to the i-th and j-th trajectory frames; λ is the learnable decay factor in the multi-head temporal attention mechanism. Let be the physical timestamp of the i-th trajectory frame; The physical timestamp of the j-th trajectory frame; The time interval between the i-th trajectory frame and the j-th trajectory frame; This is the expansion / contraction index.

[0085] It should be noted that this time bias term is a time bias term that monotonically decays with the inter-frame time interval. The form of this time bias term originates from overdamped Langevin dynamics, has a clear physical meaning, and naturally supports arbitrary floating-point time intervals. According to the above calculation formula (3), it can be seen that the time bias term only depends on the timestamp difference between the physical timestamp of the first trajectory frame and the physical timestamp of the second trajectory frame. Therefore, the initial trajectory prediction model can handle non-uniform time intervals and can generate trajectories according to a specified time resolution during inference.

[0086] Step F15: Calculate the time attention coefficient for each identical atom based on the time bias term and each of the query vectors; Specifically, the formula for calculating the time attention coefficient is as follows:

[0087] in, Let be the temporal attention coefficient of the i-th trajectory frame to the j-th trajectory frame; i and j both represent the indices of the trajectory frames. To perform summation over all frames; d represents the dimensions of the query vector and the key vector; Attention score based on content similarity; For the time offset terms corresponding to the i-th trajectory frame and the j-th trajectory frame; It should be noted that... Add attention score based on content similarity Afterwards, the time attention coefficient effectively includes The time-dependent attenuation term is used to make trajectory frames with closer time intervals more correlated, while avoiding the complete discarding of trajectory frame information with longer time intervals.

[0088] Step F16: Based on the time attention coefficient of each atom and each value vector, obtain the time attention result for each of the same atoms.

[0089] Specifically, the formula for calculating the time attention result is as follows:

[0090] in, The temporal attention result for the k-th atom in the i-th trajectory frame represents the value vector of the same atom k in different frames. Perform a weighted summation; The temporal attention coefficient between the i-th trajectory frame and the j-th trajectory frame; Let be the value vector of the k-th atom in the j-th trajectory frame.

[0091] All trajectory frames are processed through a time attention mechanism, thereby avoiding lossy compression of the generated trajectory frame information and preserving all the details of each generated trajectory frame for denoising the next target frame.

[0092] Step D14: Fuse the spatial attention result of each atom and the temporal attention result of each of the same atoms to obtain the fused potential representation of each trajectory frame; Specifically, the spatial attention result of the k-th atom in the i-th trajectory frame. The time attention result of the k-th atom in the i-th trajectory frame The fusion process is performed to obtain the atomic-level latent vector of the k-th atom in the i-th trajectory frame after fusion. When fusing temporal attention results and spatial attention results, a summation-join approach can be used. The corresponding summation-join calculation formula is shown below:

[0093] Based on this, a fused atomic-level latent vector representation of the i-th trajectory frame is constructed. .

[0094] Step D15: The fused latent representation of each trajectory frame is decoded by the diffusion decoder to obtain the predicted coordinates of each trajectory frame.

[0095] Specifically, the fused latent representations of each trajectory frame are input into the diffusion decoder for decoding to obtain the predicted coordinates of each trajectory frame. .

[0096] For example, the fused atomic-level latent representation of the i-th trajectory frame The temporal and spatial attention mechanisms in the input diffusion decoder are used for decoding to obtain decoded features; these features are then processed through the decoding layer of the diffusion decoder to obtain the denoised all-atom coordinates of the i-th trajectory frame. (The denoised all-atom coordinates are the predicted coordinates). The spatial attention mechanism of this diffusion decoder can learn the bond connections, spatial proximity, residue interactions, and ligand binding relationships between atoms within the same trajectory frame, thereby improving the prediction accuracy of the trained spatial diffusion model.

[0097] Step S50: Train the initial trajectory prediction model based on the predicted coordinates to obtain the target trajectory prediction model.

[0098] Specifically, the initial trajectory prediction model is trained based on the predicted coordinates of each trajectory frame. When the loss function of the initial trajectory prediction model converges, that is, when the loss function of the initial trajectory prediction model reaches the preset loss threshold, the predicted coordinates approximate the true coordinates. The parameters of the initial trajectory prediction model at the convergence point are fixed, and the initial trajectory prediction model at this point is used as the target trajectory prediction model.

[0099] In one feasible implementation, step S50 may include steps G11 to G13: Step G11: Obtain the true coordinates of the biomacromolecule trajectory; Step G12: Calculate the prediction error between the predicted coordinates and the actual coordinates of the biomolecule using a preset loss function; Step G13: Optimize the loss function using a preset diffusion model technique based on the prediction error. When the loss function converges, the target trajectory prediction model is obtained.

[0100] Specifically, firstly, for each trajectory frame of the biomacromolecule trajectory, its corresponding true coordinates, i.e., the three-dimensional coordinates of all atoms, are obtained. Then, the error between the predicted coordinates and the true coordinates of each trajectory frame is calculated using a loss function, such as MSE, to obtain the coordinate prediction error for each trajectory frame. Based on the obtained coordinate prediction error, the loss function is optimized using a diffusion model technique. After the loss function converges, the parameters of the initial trajectory prediction model at convergence are fixed, and this initial trajectory prediction model is used as the target trajectory prediction model. The diffusion model technique can be EDM, etc.

[0101] By implementing the first embodiment of this application, the physical timestamps of biomolecular trajectories can be introduced during the training and generation of the target trajectory prediction model, enabling the target trajectory prediction model to learn temporal dimension information. Simultaneously, the introduction of noisy biomolecular trajectories, preset noise scales, and refined representations of biomolecular molecules during the training and generation of the target trajectory prediction model allows the target trajectory prediction model to learn atomic structure dimension information of biomolecular molecules. This enables the final generated target trajectory prediction model to directly characterize the continuous dynamic behavior of biomolecular evolution over time. Furthermore, by setting a continuous temporal attention bias mechanism inspired by Langevin dynamics in the temporal attention mechanism, it can not only process all trajectory frames of noisy biomolecular trajectories simultaneously but also handle trajectory generation at arbitrary non-uniform time intervals. This solves the problem that existing machine learning models, which only use the atomic structure of biomolecular molecules as input and lack temporal dimension information, can only output static structures or conformational distributions, making it difficult to characterize the continuous dynamic behavior of biomolecular evolution over time.

[0102] Furthermore, since the initial trajectory prediction model is based on the Protenix model and transferred during its construction, the initial trajectory prediction model already possesses the basic functionality of the AlphaFold 3 architecture after its construction is completed. This significantly reduces the training computational power required to train the initial trajectory prediction model compared to training an AlphaFold 3-level geometric inference network from scratch.

[0103] Based on the first embodiment of this application, in the second embodiment of this application, the same or similar content as the first embodiment can be referred to the above description, and will not be repeated hereafter.

[0104] Based on the first embodiment described above, please refer to... Figure 2 This application proposes a trajectory prediction method that predicts the trajectories of biomacromolecules based on the target trajectory prediction model constructed in the first embodiment of this application. This addresses the problems of low timeliness and poor transferability in existing technologies where biomacromolecule trajectory prediction can only be achieved through molecular dynamics (MD) simulation.

[0105] The trajectory prediction method includes steps S01 to S06: Step S01: Pre-set the initial trajectory and initial trajectory information of a known biomacromolecule, wherein the initial trajectory includes a pre-set condition frame and a target frame to be generated, and the initial trajectory information includes the coordinates of the condition frame, the number of target frames to be generated, and the physical timestamp of the initial trajectory. Specifically, before trajectory prediction, it is necessary to pre-set the initial trajectory of a known biomolecule and its initial trajectory information. When setting the initial trajectory of a known biomolecule, at least one condition frame needs to be defined within the initial trajectory, and it is necessary to determine which frames are to be generated as target frames. Condition frames are constructed based on the determined condition frames, and the set of target frames to be generated is determined based on the determined target frames. Setting the initial trajectory information requires setting the coordinates of the condition frames, the number of target frames to be generated, and the physical timestamps of each frame included in the initial trajectory. The physical timestamps of each frame included in the initial trajectory include the physical timestamps corresponding to the condition frames and the physical timestamps of each target frame to be generated. While setting the condition frames and target frames to be generated in the initial trajectory, corresponding physical timestamps are simultaneously assigned to each condition frame and each target frame to be generated.

[0106] It should be noted that the set number of target frames to be generated determines how many target frames' predicted coordinates the target trajectory prediction model needs to generate based on the conditional frames. The number of target frames to be generated determines the length of the final generated target trajectory of the biomolecule.

[0107] For example, given the first frame or the first and last two frames of a known initial trajectory of a biomacromolecule, these are considered conditional frames, and the other trajectory frames are the target frames to be generated. The initial trajectory information includes: the number of target frames to be generated. Then the physical timestamps corresponding to the set of target frames to be generated, constructed from all target frames to be generated, can be represented as { },in, Let be the physical timestamp corresponding to the i-th target frame to be generated; the physical timestamps in the set of physical timestamps corresponding to the set of target frames to be generated are evenly distributed.

[0108] When the number of condition frames is T, the set of physical timestamps corresponding to the condition frames can be represented as: { },in, Let be the physical timestamp corresponding to the i-th conditional frame. The physical timestamps in the set of physical timestamps corresponding to the conditional frames are evenly distributed.

[0109] The set of coordinates corresponding to a condition frame can be represented as: { },in, This represents the coordinates of the i-th conditional frame. The complete physical timestamps of the initial trajectory can be obtained from the set of physical timestamps corresponding to the conditional frames and the set of physical timestamps corresponding to the target frames to be generated.

[0110] Step S02: Obtain the amino acid sequence of the known biological macromolecule protein and the ligand small molecule that binds to the protein; Specifically, the amino acid sequences of known biological macromolecules and proteins and the ligands that bind to said proteins are obtained from publicly available MD databases.

[0111] Step S03: Initialize the coordinates of the target frame set to be generated according to the preset noise scale to obtain the initial coordinates of the target frame set to be generated; Specifically, the coordinates of the target frame set to be generated are randomly initialized by a preset noise scale to obtain the initial coordinates of the target frame set to be generated. The initial coordinates of the target frame set to be generated at this time are noisy coordinates. By generating noisy initial coordinates, the denoising ability learned by the target trajectory prediction model can be combined with atomic structure dimension information to denoise them, so as to obtain the denoised predicted coordinates.

[0112] For example, a set of noise scales is pre-set. ={ },in Follows a log-normal distribution. For the largest noise scale, Let m be the m-th noise scale. This is the minimum noise scale. For any noise scale in any set of noise scales, it can be calculated using the following formula:

[0113] in, The m-th noise scale in the set of noise scales; The largest noise scale in the set of noise scales (i.e. ); The smallest noise scale in the set of noise scales (i.e. ); The exponential scheduling coefficient is the shape exponent of the noise; The total number of iterations corresponding to the noise scale set; in this application The preferred value is 160.0. The preferred value is 0.0004; The preferred value is 7.

[0114] Based on the above maximum noise scale The atomic coordinates of the target frame set to be generated are initialized to obtain the initial coordinates of the target frame set, which are represented as follows:

[0115] in, , This represents the initial coordinates (predicted coordinates) of the target frame set to be generated. Let M be the initial coordinates (predicted coordinates) of the i-th target frame in the set of target frames to be generated, and M represent the current set of target frames to be generated based on the maximum noise scale. Add noise; The noise is random sampling noise and follows a standard Gaussian distribution.

[0116] Step S04: Based on the amino acid sequence and the ligand small molecule, obtain the refined characterization of the known biomacromolecule; Specifically, the refined characterization of known biological macromolecules is obtained by describing the ligand small molecules based on their amino acid sequences. The steps for determining the refined characterization of known biological macromolecules are consistent with step S30 of the first embodiment of this application, and will not be repeated here.

[0117] Step S05: Based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, the initial coordinates of the target frame set to be generated are iteratively denoised using the target trajectory prediction model to generate the predicted coordinates of the target frame set to be generated. Specifically, based on the acquisition of the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of known biomacromolecules, the predicted coordinates of the target frame set to be generated are iteratively generated based on the target trajectory prediction model constructed in the first embodiment of this application.

[0118] In one feasible implementation, step S05 may include steps H11~H13: Step H11: Based on the initial coordinates of the target frame set to be generated, the coordinates of the conditional frame, the first preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, the first coordinate estimate of the target frame set to be generated is obtained through the target trajectory prediction model. Step H12: The first coordinate estimate of the target frame set to be generated, the initial coordinates of the target frame to be generated, and the second preset noise scale are used to generate the predicted coordinates of the target frame set to be generated through a preset diffusion model.

[0119] Step H13: Repeatedly execute the following steps: update the initial coordinates of the target frame set to be generated based on the first predicted coordinates of the target frame set to be generated; update the first preset noise scale based on the second preset noise scale; obtain the first coordinate estimate of the target frame set to be generated through the target trajectory prediction model based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the first preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules; update the second preset noise scale based on the third preset noise scale; and generate the first predicted coordinates of the target frame to be generated by passing the first coordinate estimates of each target frame to be generated, the initial coordinates of the target frame to be generated, and the second preset noise scale through a preset diffusion model. Continue this iterative denoising step until the number of times the iterative denoising step is repeated reaches a preset iteration stop condition, thereby obtaining the predicted coordinates of the target frame set to be generated.

[0120] Specifically, according to step S03, the initial coordinates of the target frame set to be generated are obtained, i.e. .

[0121] Then, based on the noise scale set from step S03... ={ Select from} External residual noise scale { The set of target frames to be generated is denoised through M-1 iterations to generate the predicted coordinates of the set of target frames to be generated.

[0122] Then, select the noise scale corresponding to the first iteration from the remaining noise scales. As the first preset noise scale for this iteration of denoising, the noise scale corresponding to the second iteration is then selected from the remaining noise scales. This serves as the second preset noise scale for this iteration.

[0123] The initial coordinates of the target frame set to be generated First preset noise scale The physical timestamp of the initial trajectory { }, Coordinates of the condition frame { The refined characterization of known biomacromolecules is input into the target trajectory prediction model to obtain the first coordinate estimate of the set of target frames to be generated. .

[0124] The first coordinate estimation of the target frame set to be generated The initial coordinates of the target frame to be generated Second preset noise scale The first predicted coordinates of the target frame set to be generated are generated using a preset diffusion model. .

[0125] Initial coordinates of the target frame set to be generated Perform the m-th iteration of denoising to generate coordinate estimates for the target frame set to be generated. The process can be represented as: For the m-th denoising step, m = {M-1, ..., 1}; the initial coordinates of the target frame set to be generated in the current iteration step m are determined. The first preset noise scale corresponding to the target frame set to be generated in the current iteration step m. Physical timestamps of all frames { }, Coordinates of the condition frame { The refined characterization of known biomacromolecules is input into a pre-trained target trajectory prediction model to obtain the first coordinate estimate of the m-th iteration of the target frame set to be generated. .

[0126] Based on the first coordinate estimate generated in the m-th iteration of the target frame set to be generated The initial coordinates of the target frame set to be generated in the m-th iteration. The second preset noise scale corresponding to the set of target frames to be generated in the previous iteration step (m-1) in the current iteration step. The predicted coordinates of the target frame set to be generated in the (m-1)th iteration are obtained using the EDM diffusion model technique. .

[0127] Repeat the process of iterating the predicted coordinates based on the (m-1)th iteration of the target frame set to be generated. Update the initial coordinates of the target frame set to be generated in the m-th iteration. (Step H11) According to the second preset noise scale Update the first preset noise scale ; Based on the initial coordinates of the target frame set to be generated The coordinates of the condition frame { }, the first preset noise scale The physical timestamp of the initial trajectory { The refined characterization of the known biomacromolecules, through the target trajectory prediction model, yields the first coordinate estimate of the set of target frames to be generated. (Repeat step H11, updating the first preset noise scale according to the second preset noise scale during execution, so as to update the first noise scale corresponding to the iteration step). According to the third preset noise scale Update the second preset noise scale Based on the first coordinate estimate generated in the m-th iteration of the target frame set to be generated. The initial coordinates of the target frame set to be generated in the (m-1)th iteration. The second preset noise scale corresponding to the set of target frames to be generated in the previous iteration step (m-2) in the current iteration step. The predicted coordinates of the target frame set to be generated in the (m-2)th iteration are obtained using the EDM diffusion model technique. (Repeat step H12, updating the second preset noise scale according to the third preset noise scale during execution). By repeatedly executing the above operation, and determining whether the current iteration denoising step has reached the preset iteration stop condition before each repetition, if the preset iteration stop condition is reached, stop the iterative denoising of the initial coordinates of the target frame set to be generated, and change the predicted coordinates of the target frame set to be generated when the preset iteration stop condition is reached. .in, In this application, the preset iteration stopping condition can be either the third preset noise scale being equal to the minimum noise scale of the preset noise scale, or the number of iterations being equal to the number of noise scales minus 2.

[0128] Step S06: Based on the coordinates of the conditional frame and the predicted coordinates of the set of target frames to be generated, generate the target trajectory of the known biomacromolecule.

[0129] Specifically, based on the order of each target frame in the initial trajectory within the conditional frame and the set of target frames to be generated, the coordinates of the conditional frame and the predicted coordinates of the set of target frames to be generated are combined to generate the target trajectory of the known biomacromolecule. That is, the coordinates of the conditional frame { } and the predicted coordinates of the target frame set to be generated By sequentially assembling the sequences, the target trajectories of known biological macromolecules can be obtained. }

[0130] By predicting the coordinates of each target frame to be generated within the conditional frame based on the true coordinates of the conditional frame, the target trajectory of a known biomolecule can be predicted. The generated target trajectory includes the true coordinates of the conditional frame and the predicted coordinates of the target frame to be generated, and can be used to analyze protein conformational changes, ligand binding or dissociation pathways, allosteric interactions, and kinetic transitions between different conformational states.

[0131] By implementing the second embodiment of this application, a target trajectory for a known biomacromolecule can be achieved by pre-configuring an initial trajectory containing a conditional frame and a target frame to be generated, initial trajectory information, the amino acid sequence of a known biomacromolecule protein, and a small ligand molecule binding to the protein. The prediction of this target trajectory incorporates temporal and atomic structure information, ensuring that the generated target trajectory for the known biomacromolecule includes temporally correlated conditional frame coordinates and predicted coordinates of each target frame to be generated. This allows the target trajectory to be used to analyze protein conformational changes, ligand binding or dissociation pathways, allosteric interactions, and dynamic transitions between different conformational states. Furthermore, the target trajectory generated based on a diffusion model transforms the molecular dynamics simulation process from traditional molecular dynamics (MD) simulation to machine learning simulation. Leveraging the high-speed processing and reusable network parameters of machine learning, it outperforms traditional MD simulation in terms of the timeliness and transferability of target trajectory generation.

[0132] Based on the first and second embodiments described above, this application proposes a third embodiment, which is a supplement to another feasible implementation of steps S05 and S06.

[0133] Reference Figure 3 In the third embodiment of this application, steps S05 and S06 can be implemented by steps I11 to I16.

[0134] Step I11: If the difference in physical timestamps of the target frame set to be generated is greater than a preset threshold, the physical timestamps of the target frame set to be generated are fragmented to obtain multiple fragmented target frames and multiple fragmented physical timestamps of the fragmented target frames. Specifically, the physical timestamp difference of the target frames to be generated is determined based on the physical timestamp of the set of target frames to be generated set in step S01.

[0135] For example, suppose the physical timestamp of the target frame set to be generated is { }, the difference in physical timestamps of the target frame set to be generated ( When the physical timestamp of the target frame set to be generated exceeds a preset threshold (e.g., exceeding 100ns), the physical timestamp of the target frame set to be generated is fragmented to obtain multiple fragmented target frames and multiple fragmented physical timestamps of the target frames. The physical timestamp of the fragmented target frame set to be generated can be represented as { }.in, .

[0136] Step I12: For each segmented target frame, based on the segmented target frame and its segmented physical timestamp, the sparse predicted coordinates of each segmented target frame are obtained through the target trajectory prediction model. Specifically, each segmented target frame and its segmented physical timestamp are input into the target trajectory prediction model to obtain the sparse predicted coordinates of each segmented target frame. Among them, the sparse predicted coordinates of the target frame in the segmentation. It can be represented as: ,in, These are the sparse predicted coordinates of the k-th fragment of the target frame.

[0137] In one feasible implementation, step I12 may include steps J11~J12: Step J11: For any segmented target frame, use the segmented physical timestamp of the segmented target frame as the physical timestamp of the initial trajectory, use the segmented target frame as the initial trajectory, and return to execute the step of generating the predicted coordinates of the target frame set to be generated through the target trajectory prediction model based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, until the predicted coordinates of all segmented target frames are generated; Specifically, when inputting the fragmented target frame and its fragmented physical timestamp into the target trajectory prediction model, the fragmented physical timestamp of each target frame is used as the physical timestamp of the initial trajectory. After using the fragmented target frame as the initial trajectory, the process returns to step S05, which involves generating the predicted coordinates of the target frame set to be generated. The predicted coordinates of the target frame set to be generated are then used as the predicted coordinates of the current fragmented target frame. This process continues until the target trajectory prediction model outputs the predicted coordinates of all fragmented target frames.

[0138] Step J12: Obtain the sparse prediction coordinates of each of the target fragments based on the predicted coordinates of all the target fragments.

[0139] Step I13: For each intermediate target frame between adjacent fragmented target frames, set a target condition frame and an intermediate target frame to be generated in each intermediate target frame, and set a physical timestamp for the intermediate target frame. Each adjacent fragmented target frame is used as a target condition frame, the sparse prediction coordinates corresponding to the adjacent fragmented target frames are used as the coordinates of the target condition frame, and the other target frames in each intermediate target frame except the target condition frame are set as intermediate target frames to be generated. Specifically, for adjacent target fragments and Fine-grained interpolation is performed on k intermediate target frames between them. (The sentence is incomplete and ends abruptly.) and Set as the target condition frame, and and The corresponding sparse prediction coordinates are set as the coordinates of the target conditional frame; and The k intermediate target frames between are used as intermediate target frames to be generated, and the timestamps corresponding to the intermediate target frames to be generated are set to { }

[0140] Step I14: Based on each target condition frame, the coordinates of each target condition frame, each intermediate target frame to be generated, and the physical timestamp of the intermediate target frame, the predicted coordinates of each intermediate target frame to be generated are generated using the target trajectory prediction model. Specifically, each target condition frame, the coordinates of each target condition frame, each intermediate target frame to be generated, and the physical timestamp of the intermediate target frame belong to the target trajectory prediction model, which outputs the predicted coordinates of each intermediate target frame to be generated.

[0141] For example, for a set of adjacent target fragments, when the adjacent target fragments are... and , and The corresponding sparse prediction coordinates are set to the coordinates of the target conditional frame. and ,based on and The intermediate target frames to be generated are determined by the k intermediate target frames between them, and the timestamps corresponding to the intermediate target frames to be generated are { After inputting the target trajectory prediction model, the target trajectory prediction model outputs the predicted coordinates corresponding to the intermediate target frames to be generated. .

[0142] In one feasible implementation, step I14 may include K11~K12: Step K11: For any intermediate target frame, initialize the coordinates of the intermediate target frame to be generated based on the preset noise scale to obtain the initial coordinates of each intermediate target frame to be generated. Specifically, the coordinates of the intermediate target frame to be generated are initialized based on the above formula (8). The corresponding coordinate initialization logic is described in the implementation description of step S03 of this application, and will not be repeated here.

[0143] Step K12: Based on the initial coordinates of the intermediate target frame to be generated, the coordinates of the target condition frame, the preset noise scale, the physical timestamp of the intermediate target frame to be generated, and the refined characterization of the known biomacromolecule, the predicted coordinates of each remaining intermediate target frame to be generated are generated through the target trajectory prediction model.

[0144] Specifically, the initial coordinates of the intermediate target frame to be generated, the coordinates of the target condition frame, the preset noise scale, the physical timestamp of the intermediate target frame to be generated, and the refined characterization of the known biomacromolecule are input into the target trajectory prediction model, and the predicted coordinates of the remaining intermediate target frames to be generated are generated through the target trajectory prediction model.

[0145] Step I15: Based on the coordinates of each target condition frame and the predicted coordinates of each intermediate target frame, generate the coordinates of each intermediate target frame to be generated; Specifically, for each group of adjacent target fragments, the coordinates of the target condition frame and the predicted coordinates of each intermediate target frame are combined to obtain the coordinates of the intermediate target frame to be generated for the current adjacent target fragment.

[0146] Step I16: Based on the coordinates of all intermediate frames to be generated, obtain the target trajectory of the known biomacromolecule.

[0147] Specifically, the coordinates of each intermediate frame to be generated are connected according to the physical timestamps corresponding to the coordinates of all intermediate frames to be generated, so as to obtain the target trajectory of the known biomolecule.

[0148] By implementing steps I11-I16 above, a sparse keyframe sequence is generated at long time intervals through autoregressive prediction (corresponding to the target frame generation stages of each segment, i.e., steps I11-I12). Then, using these keyframes as anchor points at both ends, all the intermediate fine frames are generated in parallel through interpolation (corresponding to the intermediate target frame generation stage, i.e., steps I15-I16). The same target trajectory prediction model is used in both stages, differing only in the preset noise scale configuration for each frame. This target trajectory prediction model determines the global direction of the trajectory with the fewest autoregressive steps, and then fills in the local details with parallel interpolation constrained by anchor points, thereby significantly mitigating the error accumulation in autoregressive generation. Using the above strategy, it is possible to stably generate all-atom trajectories, maintain conformational stability over long time, avoid significant divergence, and output trajectories at any specified temporal resolution.

[0149] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the trajectory prediction model construction method and trajectory prediction method of this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0150] This application also provides a trajectory prediction model construction device, please refer to... Figure 4 The trajectory prediction model construction device includes: a training data preprocessing module 10, a prediction coordinate generation module 20, and a model training module 30; The training data preprocessing module 10 is used to acquire the biomacromolecule trajectory, the physical timestamp of the biomacromolecule trajectory, the amino acid sequence of the biomacromolecule protein, and the ligand small molecule bound to the protein; add noise to the biomacromolecule trajectory based on a preset noise scale to generate a noisy biomacromolecule trajectory; and obtain a refined characterization of the biomacromolecule based on the amino acid sequence and the ligand small molecule. The predicted coordinate generation module 20 is used to predict the predicted coordinates of the biomacromolecule based on the noisy biomacromolecule trajectory, the preset noise scale, the physical timestamp of the biomacromolecule trajectory, and the refined characterization of the biomacromolecule through an initial trajectory prediction model, wherein the initial trajectory prediction model includes a diffusion encoder and a diffusion decoder. The model training module 30 is used to train the initial trajectory prediction model based on the predicted coordinates to obtain the target trajectory prediction model.

[0151] The trajectory prediction model construction apparatus provided in this application, employing the trajectory prediction model construction method in the above embodiments, can solve the technical problem that existing technologies struggle to generate biomolecular dynamic trajectories using machine learning models. Compared with existing technologies, the beneficial effects of the trajectory prediction model construction apparatus provided in this application are the same as those of the trajectory prediction model construction method provided in the above embodiments, and other technical features in the trajectory prediction model construction apparatus are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0152] This application provides a trajectory prediction model building device, which includes: at least one processor; and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the trajectory prediction model building method in Embodiment 1 above.

[0153] This application also provides a trajectory prediction device, please refer to... Figure 5 The trajectory prediction device includes: a data preprocessing module 40, a coordinate prediction module 50, and a trajectory prediction module 60; The data preprocessing module 40 is used to pre-set the initial trajectory and initial trajectory information of a known biomacromolecule, wherein the initial trajectory includes a pre-set condition frame and a set of target frames to be generated, and the initial trajectory information includes the coordinates of the condition frame, the number of target frames to be generated, and the physical timestamp of the initial trajectory; acquire the amino acid sequence of the known biomacromolecule protein and the ligand small molecule that binds to the protein; initialize the coordinates of the set of target frames to be generated according to a preset noise scale to obtain the initial coordinates of the set of target frames to be generated; and obtain the refined characterization of the known biomacromolecule based on the amino acid sequence and the ligand small molecule. The coordinate prediction module 50 is used to iteratively denoise the initial coordinates of the target frame set to be generated based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, through the target trajectory prediction model, to generate the predicted coordinates of the target frame set to be generated. The trajectory prediction module 60 is used to generate the target trajectory of the known biomacromolecule based on the coordinates of the condition frame and the predicted coordinates of each of the target frames to be generated.

[0154] The trajectory prediction device provided in this application, employing the trajectory prediction method described in the above embodiments, can solve the technical problems of low timeliness and poor transferability of existing MD simulations of biomolecular dynamics trajectories. Compared with the prior art, the beneficial effects of the trajectory prediction method provided in this application in addressing the low timeliness and poor transferability of existing MD simulations of biomolecular dynamics trajectories are the same as those of the trajectory prediction method provided in the above embodiments, and other technical features in the trajectory prediction device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0155] The following is for reference. Figure 6 The diagram illustrates a structural schematic suitable for implementing the trajectory prediction model building device of the embodiments of this application. The trajectory prediction model building device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The trajectory prediction model building device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0156] like Figure 6 As shown, the trajectory prediction model building device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 allows the trajectory prediction model building device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a trajectory prediction model building device with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented alternatively.

[0157] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0158] The trajectory prediction model construction device provided in this application, employing the trajectory prediction model construction method in the above embodiments, can solve the technical problem that existing technologies struggle to generate biomolecular dynamic trajectories using machine learning models. Compared with existing technologies, the beneficial effects of the trajectory prediction model construction device provided in this application are the same as those of the trajectory prediction model construction method provided in the above embodiments, and other technical features in this trajectory prediction model construction device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0159] The following is for reference. Figure 7 The diagram illustrates a structural schematic suitable for implementing the biomolecular trajectory prediction device of the embodiments of this application. The biomolecular trajectory prediction device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 7 The illustrated biomolecular trajectory prediction device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0160] like Figure 7 As shown, the biomolecular trajectory prediction device may include a processing unit 2001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 2002 or a program loaded from a storage device 2003 into a random access memory (RAM) 2004. The RAM 1004 also stores various programs and data required for the operation of the device. The processing unit 2001, ROM 1002, and RAM 1004 are interconnected via a bus 2005. An input / output (I / O) interface 2006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 2006: input devices 2007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 2008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 2003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 2009. Communication device 2009 allows the biomacromolecule trajectory prediction device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows biomacromolecule trajectory prediction devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0161] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 2003, or installed from ROM 2002. When the computer program is executed by processing device 2001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0162] The biomolecular trajectory prediction device provided in this application, employing the trajectory prediction method described in the above embodiments, can solve the technical problems of low timeliness and poor transferability in existing MD simulations of biomolecular dynamic trajectories. Compared with the prior art, the beneficial effects of the biomolecular trajectory prediction device provided in this application are the same as those of the trajectory prediction method provided in the above embodiments, and other technical features of this biomolecular trajectory prediction device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0163] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0164] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0165] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the trajectory prediction model construction method and trajectory prediction method in the above embodiments.

[0166] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0167] The aforementioned computer-readable storage medium may be included in the trajectory prediction model building device or the trajectory prediction device itself; or it may exist independently and not be assembled into the trajectory prediction model building device.

[0168] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the trajectory prediction model building device, cause the trajectory prediction model building device to perform a trajectory prediction model building method. When the aforementioned one or more programs are executed by the trajectory prediction device, the trajectory prediction device also performs a trajectory prediction method.

[0169] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0170] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0171] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0172] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-described trajectory prediction model construction method and trajectory prediction method. This addresses the technical problems of existing technologies, such as the difficulty in generating biomolecular dynamic trajectories using machine learning models and the low timeliness and transferability of existing MD simulations of biomolecular dynamic trajectories. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the trajectory prediction model construction method and trajectory prediction method provided in the above embodiments, and will not be repeated here.

[0173] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the trajectory prediction model construction method and the trajectory prediction method as described above.

[0174] The computer program product provided in this application can solve the technical problems of existing technologies, such as the difficulty in generating biomolecular dynamic trajectories using machine learning models and the low timeliness and transferability of existing MD simulations of biomolecular dynamic trajectories. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the trajectory prediction model construction method and trajectory prediction method provided in the above embodiments, and will not be repeated here.

[0175] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for constructing a trajectory prediction model, characterized in that, The trajectory prediction model construction method includes: Obtain the trajectory of a biomacromolecule, the physical timestamp of the biomacromolecule trajectory, the amino acid sequence of the biomacromolecule protein, and the ligand small molecule that binds to the protein; Noise is added to the biomacromolecule trajectory based on a preset noise scale to generate a noisy biomacromolecule trajectory. The refined characterization of the biomacromolecule was obtained based on the amino acid sequence and the ligand small molecule. Based on the noisy biomacromolecule trajectory, the preset noise scale, the physical timestamp of the biomacromolecule trajectory, and the refined characterization of the biomacromolecule, the predicted coordinates of the biomacromolecule are obtained by predicting using an initial trajectory prediction model, wherein the initial trajectory prediction model includes a diffusion encoder and a diffusion decoder. The initial trajectory prediction model is trained based on the predicted coordinates to obtain the target trajectory prediction model.

2. The trajectory prediction model construction method as described in claim 1, characterized in that, The step of adding noise to the biomacromolecule trajectory based on a preset noise scale to generate a noisy biomacromolecule trajectory includes: Set a conditional frame for the biomacromolecule trajectory, and set the other trajectory frames in the biomacromolecule trajectory other than the conditional frame as the target frame; Based on the preset noise scale, noise is added to the conditional frame and each of the target frames to obtain a noisy conditional frame and each noisy target frame. Based on the noisy conditional frames and each of the noisy target frames, a noisy biomacromolecule trajectory is generated.

3. The trajectory prediction model construction method as described in claim 1, characterized in that, The refining and characterization of the biomacromolecule obtained based on the amino acid sequence and the ligand small molecule includes: Structural templates for each type of amino acid, structural templates for the ligand small molecules, and atomic type characteristics of the biomacromolecules are obtained from a pre-set database. The amino acid sequence is processed to obtain the sequence alignment of the amino acid sequence; The atomic topology of the ligand small molecule is obtained based on the ligand small molecule; The amino acid sequence, the atomic topology, the sequence alignment, the structural template of each type of amino acid, the structural template of the ligand small molecule, and the atomic type features of the biomacromolecule are passed through a preset feature encoder to obtain a single representation of each word. Based on the single representation of each lexical unit, a refined representation of the biological macromolecule is obtained.

4. The trajectory prediction model construction method as described in claim 3, characterized in that, The refined characterization of the biological macromolecule based on the single characterization of each lexical element includes: Pair all lexical units together to obtain several lexical pairs; After linear transformation, the pair representations of each lexical pair are obtained; Triangular updates and attention interactions are performed on the single representations of all lexical units and the pair representations of all lexical pairs to obtain the refined single representations and refined pair representations of the biomacromolecules. The refined characterization of the biomacromolecule is obtained based on the single characterization of the refined biomacromolecule and the pair characterization of the refined biomacromolecule.

5. The trajectory prediction model construction method as described in claim 1, characterized in that, The prediction of the biomacromolecule's predicted coordinates, based on the noisy biomacromolecule trajectory, the preset noise scale, the physical timestamp of the biomacromolecule trajectory, and the refined characterization of the biomacromolecule, is performed using an initial trajectory prediction model. This includes: The noisy biomacromolecule trajectory, the preset noise scale, the noisy biomacromolecule trajectory, the refined characterization of the biomacromolecule, and the atomic type features of the biomacromolecule are passed through the diffusion encoder to obtain the latent representation of each trajectory frame. One trajectory frame contains multiple atoms, and the latent representation of one trajectory frame is composed of the latent vectors of multiple atoms. For each atom in each trajectory frame, the spatial attention result for each atom is obtained through the spatial attention mechanism of the diffusion encoder based on the latent vector of each atom; For each identical atom in any two trajectory frames, the temporal attention result for each identical atom is obtained through the temporal attention mechanism of the diffusion encoder based on the latent vector of each identical atom and the physical timestamp of the two trajectory frames. The spatial attention result of each atom and the temporal attention result of the same atom are fused to obtain the fused potential representation of each trajectory frame; The fused latent representation of each trajectory frame is decoded by the diffusion decoder to obtain the predicted coordinates of each trajectory frame.

6. The trajectory prediction model construction method as described in claim 5, wherein for each atom of each trajectory frame, obtaining the spatial attention result of each atom based on the latent vector of each atom through the spatial attention mechanism of the diffusion encoder includes: A linear transformation is performed on the potential vector of each first atom in each trajectory frame to obtain the query vector of each trajectory frame, wherein the first atom is any atom in each trajectory frame; A linear transformation is performed on the potential vector of the second atom in each trajectory frame to obtain the bond vector and value vector of each trajectory frame, wherein the second atom is any atom other than the first atom in each trajectory frame; A linear transformation was performed on the refined characterization of the biomacromolecules to obtain the spatial bias term; The spatial attention coefficient of each atom is calculated based on the spatial bias term, the query vector of each trajectory frame, the first atom of each trajectory frame, the second atom of each trajectory frame, and the third atom associated with the first atom. The spatial attention result for each atom is obtained based on the spatial attention coefficient of each atom and the value vector of each trajectory frame.

7. The trajectory prediction model construction method as described in claim 5, wherein for each identical atom in any two trajectory frames, based on the latent vector of each identical atom and the physical timestamps of the two trajectory frames, the temporal attention result of each identical atom is obtained through the temporal attention mechanism of the diffusion encoder, comprising: From the noisy biomacromolecule trajectories, select all first trajectory frames and second trajectory frames that have the same atom, wherein the number of the same atom is at least one; A linear transformation is performed on the potential vector of each identical atom in each of the first trajectory frames to obtain the query vector of each identical atom in the first trajectory frame; A linear transformation is performed on the latent vector of each identical atom in each of the second trajectory frames to obtain the bond vector and value vector of each identical atom in the second trajectory frame; A time bias term is generated based on the time attention mechanism, the physical timestamp of the first trajectory frame, and the physical timestamp of the second trajectory frame; Calculate the time attention coefficient for each identical atom based on the time bias term and each of the query vectors; The time attention result for each of the same atoms is obtained based on the time attention coefficient of each atom and each value vector.

8. The trajectory prediction model construction method according to any one of claims 1-7, wherein training the initial trajectory prediction model based on the predicted coordinates to obtain the target trajectory prediction model includes: Obtain the true coordinates of the biomolecule's trajectory; The prediction error between the predicted coordinates and the actual coordinates of the biomolecule is calculated using a preset loss function. The loss function is optimized using a preset diffusion model technique based on the prediction error, and the target trajectory prediction model is obtained when the loss function converges.

9. A trajectory prediction method, characterized in that, A target trajectory prediction model is constructed using the trajectory prediction model construction method as described in any one of claims 1-8, wherein the trajectory prediction method includes: The initial trajectory and initial trajectory information of a known biomacromolecule are preset, wherein the initial trajectory includes a preset condition frame and a set of target frames to be generated, and the initial trajectory information includes the coordinates of the condition frame, the number of the target frames to be generated, and the physical timestamp of the initial trajectory; Obtain the amino acid sequence of the known biological macromolecule protein and the ligand small molecule that binds to the protein; The coordinates of the target frame set to be generated are initialized according to a preset noise scale to obtain the initial coordinates of the target frame set to be generated; The refined characterization of the known biomacromolecules was obtained based on the amino acid sequence and the ligand small molecule. Based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, the target trajectory prediction model iteratively denoises the initial coordinates of the target frame set to be generated, thereby generating the predicted coordinates of the target frame set to be generated. Based on the coordinates of the conditional frame and the predicted coordinates of the set of target frames to be generated, the target trajectory of the known biomacromolecule is generated.

10. The trajectory prediction method as described in claim 9, characterized in that, The step of iteratively denoising the initial coordinates of the target frame set to be generated using the target trajectory prediction model based on the initial coordinates of the target frame set to be generated, the coordinates of the conditional frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, to generate the predicted coordinates of the target frame set to be generated, includes: Based on the initial coordinates of the target frame set to be generated, the coordinates of the conditional frame, the first preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, the first coordinate estimate of the target frame set to be generated is obtained through the target trajectory prediction model. The first coordinate estimate of the target frame set to be generated, the initial coordinates of the target frame to be generated, and the second preset noise scale are used to generate the first predicted coordinates of the target frame set to be generated through a preset diffusion model; The process involves repeatedly updating the initial coordinates of the target frame set to be generated based on the first predicted coordinates, and updating the first preset noise scale based on the second preset noise scale. Then, using the target trajectory prediction model, the initial coordinates of the target frame set to be generated, the coordinates of the conditional frame, the first preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules are used to obtain the first coordinate estimate of the target frame set to be generated. The second preset noise scale is then updated based on the third preset noise scale. Finally, an iterative denoising step is performed, using the first coordinate estimates of each target frame to be generated, the initial coordinates of the target frame to be generated, and the second preset noise scale, through a preset diffusion model to generate the first predicted coordinates of the target frame to be generated. This process continues until the number of iterations of the iterative denoising step reaches a preset iteration stop condition, thus obtaining the predicted coordinates of the target frame set to be generated.

11. The trajectory prediction method according to claim 9, characterized in that, Based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, the target trajectory prediction model iteratively denoises the initial coordinates of the target frame set to be generated, thereby generating the predicted coordinates of the target frame set to be generated. Based on the coordinates of the conditional frame and the predicted coordinates of the set of target frames to be generated, the target trajectory of the known biomacromolecule is generated, including: If the difference in physical timestamps of the target frame set to be generated is greater than a preset threshold, the physical timestamps of the target frames to be generated are fragmented to obtain multiple fragmented target frames and multiple fragmented physical timestamps of the fragmented target frames. For each segmented target frame, based on the segmented target frame and its segmented physical timestamp, the sparse predicted coordinates of each segmented target frame are obtained through the target trajectory prediction model. For each intermediate target frame between adjacent fragmented target frames, a target condition frame and an intermediate target frame to be generated are set in each intermediate target frame, and a physical timestamp is set for the intermediate target frame. Each adjacent fragmented target frame is used as the target condition frame, the sparse prediction coordinates corresponding to the adjacent fragmented target frames are used as the coordinates of the target condition frame, and the other target frames in each intermediate target frame except the target condition frame are set as intermediate target frames to be generated. Based on each target condition frame, the coordinates of each target condition frame, each intermediate target frame to be generated, and the physical timestamp of the intermediate target frame, the predicted coordinates of each intermediate target frame to be generated are generated by the target trajectory prediction model. Based on the coordinates of each target condition frame and the predicted coordinates of each intermediate target frame, the coordinates of each intermediate target frame to be generated are generated. Based on the coordinates of all intermediate frames to be generated, the target trajectory of the known biomacromolecule is obtained.

12. The trajectory prediction method according to claim 11, characterized in that, For each segmented target frame, the sparse predicted coordinates of each segmented target frame are obtained through the target trajectory prediction model based on the segmented target frame and its segmented physical timestamp, including: For any segmented target frame, the segmented physical timestamp of the segmented target frame is used as the physical timestamp of the initial trajectory, the segmented target frame is used as the initial trajectory, and the process returns to execute the step of generating the predicted coordinates of the target frame set to be generated through the target trajectory prediction model based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, until the predicted coordinates of all segmented target frames are generated; The sparse predicted coordinates of each of the target fragments are obtained based on the predicted coordinates of all the target fragments.

13. The trajectory prediction method according to claim 12, characterized in that, The step of generating predicted coordinates for each intermediate target frame to be generated using the target trajectory prediction model based on each target condition frame, the coordinates of each target condition frame, each intermediate target frame to be generated, and the physical timestamp of each intermediate target frame includes: For any intermediate target frame, the coordinates of the intermediate target frame to be generated are initialized based on the preset noise scale to obtain the initial coordinates of each intermediate target frame to be generated; Based on the initial coordinates of the intermediate target frame to be generated, the coordinates of the target condition frame, the preset noise scale, the physical timestamp of the intermediate target frame to be generated, and the refined characterization of the known biomacromolecule, the predicted coordinates of each remaining intermediate target frame to be generated are generated through the target trajectory prediction model.

14. A trajectory prediction model construction device, characterized in that, The trajectory prediction model building device includes: a training data preprocessing module, a prediction coordinate generation module, and a model training module; The training data preprocessing module is used to acquire biomacromolecule trajectories, physical timestamps of the biomacromolecule trajectories, amino acid sequences of the biomacromolecule proteins, and small ligand molecules bound to the proteins; add noise to the biomacromolecule trajectories based on a preset noise scale to generate noisy biomacromolecule trajectories; and obtain refined characterization of the biomacromolecules based on the amino acid sequences and the small ligand molecules. The predicted coordinate generation module is used to predict the predicted coordinates of the biomacromolecule based on the noisy biomacromolecule trajectory, the preset noise scale, the physical timestamp of the biomacromolecule trajectory, and the refined characterization of the biomacromolecule through an initial trajectory prediction model, wherein the initial trajectory prediction model includes a diffusion encoder and a diffusion decoder. The model training module is used to train the initial trajectory prediction model based on the predicted coordinates to obtain the target trajectory prediction model.

15. A trajectory prediction device, characterized in that, The trajectory prediction device includes: a data preprocessing module, a coordinate prediction module, and a trajectory prediction module; The data preprocessing module is used to pre-set the initial trajectory and initial trajectory information of a known biomacromolecule, wherein the initial trajectory includes a pre-set condition frame and a set of target frames to be generated, and the initial trajectory information includes the coordinates of the condition frame, the number of target frames to be generated, and the set of physical timestamps of the initial trajectory; acquire the amino acid sequence of the known biomacromolecule protein and the ligand small molecule that binds to the protein; initialize the coordinates of the set of target frames to be generated according to a preset noise scale to obtain the initial coordinates of the set of target frames to be generated; and obtain the refined characterization of the known biomacromolecule based on the amino acid sequence and the ligand small molecule. The coordinate prediction module is used to iteratively denoise the initial coordinates of the target frame set to be generated based on the initial coordinates of the target frame set to be generated, the coordinates of the condition frame, the preset noise scale, the physical timestamp of the initial trajectory, and the refined characterization of the known biomacromolecules, through the target trajectory prediction model, to generate the predicted coordinates of the target frame set to be generated. The trajectory prediction module is used to generate the target trajectory of the known biomacromolecule based on the coordinates of the condition frame and the predicted coordinates of the set of target frames to be generated.

16. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and one or more computer programs are stored on the storage medium. When the computer programs are executed by a processor, they implement the steps of the trajectory prediction model construction method as described in any one of claims 1 to 8, and / or implement the steps of the trajectory prediction method as described in any one of claims 9 to 13.

17. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the trajectory prediction model construction method as described in any one of claims 1 to 8, and / or implements the steps of the trajectory prediction method as described in any one of claims 9 to 13.

Images