Method, apparatus, medium, and program product for predicting protein residue motion coordinates

Abstract

The application discloses a protein residue motion coordinate prediction method, device, medium and program product, relates to the technical field of protein dynamics, and the protein residue motion coordinate prediction method comprises the following steps: acquiring a first frequency spectrum corresponding to a coordinate sequence of a protein residue, and acquiring a characteristic component in a preset frequency range from the first frequency spectrum to determine a frequency domain feature vector, wherein the preset frequency range is lower than a preset frequency threshold; inputting a structure feature vector, a sequence feature vector and the frequency domain feature vector of the protein residue into a preset latent diffusion model, performing diffusion prediction on the protein residue by the latent diffusion model, and determining a predicted frequency domain vector; converting the predicted frequency domain vector into a corresponding second frequency spectrum, and restoring the second frequency spectrum into a predicted coordinate sequence of the protein residue. The application solves the technical problems of large calculation amount and low efficiency of a traditional protein residue motion trajectory prediction scheme.

Description

Technical Field

[0001] This application relates to the field of protein dynamics technology, and in particular to a method, device, computer-readable storage medium, and computer program product for predicting the motion coordinates of protein residues. Background Technology

[0002] Currently, in the field of protein dynamics prediction, when faced with the need to model and predict the trajectory of protein residues, the traditional approach is to predict the three-dimensional coordinates of protein residues frame by frame in the time domain at each time point. However, as the number of prediction steps increases, the input and output dimensions of the model increase linearly, leading to a surge in computational load, a significant increase in inference time, low efficiency, and difficulty in meeting the needs for long-term conformational changes.

[0003] The information disclosed in this background section is only for understanding the background technology of the present application concept, and therefore may contain information that does not constitute prior art. Summary of the Invention

[0004] The main purpose of this application is to provide a method, device, medium and program product for predicting the motion coordinates of protein residues, which aims to solve the technical problems of large computational load and low efficiency in traditional protein residue motion trajectory prediction schemes.

[0005] To achieve the above objectives, this application provides a method for predicting the motion coordinates of protein residues, the method comprising:

[0006] Obtain the first spectrum corresponding to the coordinate sequence of protein residues, and extract the feature components within a preset frequency range from the first spectrum to determine the frequency domain feature vector, wherein the preset frequency range is lower than a preset frequency threshold;

[0007] The structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues are input into a preset potential diffusion model, and the potential diffusion model performs diffusion prediction on the protein residues to determine the predicted frequency domain vector.

[0008] The predicted frequency domain vector is converted into the corresponding second spectrum, and the second spectrum is restored to the predicted coordinate sequence of protein residues.

[0009] In one embodiment, the step of obtaining a first spectrum corresponding to the coordinate sequence of protein residues, and obtaining feature components within a preset frequency range from the first spectrum to determine a frequency domain feature vector includes:

[0010] A fast Fourier transform is performed on the coordinate sequence of the protein residues to obtain the first spectrum of the time-series signal corresponding to each of the three dimensions;

[0011] Extract a preset number of low-frequency components with the largest amplitude from each of the first spectra, or select low-frequency components within a preset frequency range from each of the first spectra;

[0012] The low-frequency components are concatenated into a frequency domain feature vector based on the protein residues and the three-dimensional coordinate axis order.

[0013] In one embodiment, the potential diffusion model includes a normalization module and a flow network module;

[0014] The step of inputting the structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues into a preset latent diffusion model, and having the latent diffusion model predict the diffusion of the protein residues to determine the predicted frequency domain vector includes:

[0015] After the structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues are input into a preset potential diffusion model, the frequency domain feature vector is standardized by the standardization module to obtain a standard frequency domain vector.

[0016] The standard frequency domain vector is processed by the residual block in the flow network module, and then the processed standard frequency domain vector, the sequence feature vector, and the structural feature vector are fused into the attention calculation to obtain a multi-dimensional frequency domain feature vector.

[0017] The multidimensional frequency domain feature vector is initialized based on a normal distribution, and the multidimensional frequency domain feature vector is denoised iteratively to obtain the predicted feature vector.

[0018] In one embodiment, before the step of inputting the structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues into a preset potential diffusion model, the method further includes:

[0019] To obtain the initial static structure and amino acid sequence of protein residues;

[0020] The initial static structure is encoded by a preset variational autoencoder to obtain the structural feature vector of the protein residues;

[0021] Extract the sequence feature vector corresponding to the amino acid sequence.

[0022] In one embodiment, the protein residue motion coordinate prediction method further includes:

[0023] The degree features of protein residues are obtained, and these degree features are determined as the reconstruction content target of the initial diffusion model.

[0024] The initial diffusion model is trained based on the reconstructed content target to obtain a pre-trained model;

[0025] The pre-trained model is transferred to a preset spectrum reconstruction task, and the pre-trained model is trained with the goal of minimizing the frequency domain reconstruction error to obtain a potential diffusion model.

[0026] In one embodiment, the step of converting the predicted frequency domain vector into a corresponding second spectrum and restoring the second spectrum to a predicted coordinate sequence of protein residues includes:

[0027] The predicted frequency domain vector is filled with high-frequency zeros to obtain the second spectrum;

[0028] Perform an inverse fast Fourier transform on the second spectrum to obtain the predicted coordinate sequence of protein residues.

[0029] In one embodiment, prior to the step of obtaining the first spectrum corresponding to the coordinate sequence of protein residues, the method further includes:

[0030] Acquire raw three-dimensional structural data and motion trajectory data of proteins;

[0031] Based on the motion trajectory data, the three-dimensional coordinates of each protein residue are recorded in time series form to obtain the coordinate sequence of the protein residues.

[0032] In addition, to achieve the above objectives, this application also provides a protein residue motion coordinate prediction device, which includes at least: 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 protein residue motion coordinate prediction method as described above.

[0033] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the protein residue motion coordinate prediction method described above.

[0034] In addition, to achieve the above objectives, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the protein residue motion coordinate prediction method as described above.

[0035] This application provides a method for predicting the motion coordinates of protein residues. First, a first spectrum corresponding to the coordinate sequence of protein residues is obtained. Then, feature components within a preset frequency range are extracted from the first spectrum to determine a frequency domain feature vector. Using the low-frequency frequency domain vector as the target representation of the diffusion model significantly compresses the original multidimensional time-series signal, reduces the scale of the model's input and output, and thus significantly reduces the computational overhead of training and inference. Next, the structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues are input into a preset latent diffusion model. The latent diffusion model performs diffusion prediction on the protein residues to determine the predicted frequency domain vector. Then, the predicted frequency domain vector is converted into a corresponding second spectrum, and the second spectrum is restored to the predicted coordinate sequence of the protein residues. In this technical solution, to avoid multiple iterations when restoring the time-domain trajectory, the one-time generation advantage of the latent diffusion model in image generation is utilized. It can map the frequency domain feature components to a compact latent space, allowing the diffusion model to reconstruct all multiple low-frequency coefficients in this space at once, thereby obtaining a complete frequency domain representation, i.e., the predicted frequency domain. Finally, it is restored back to the time domain to obtain the predicted coordinate sequence of the protein residues. The protein residue motion coordinate prediction method of this application not only avoids the repetitive calculation of frame-by-frame prediction and reduces the amount of computation, but also achieves rapid generation of motion trajectory coordinates by taking advantage of the fewer parameters and more concentrated expression in the latent diffusion model, thereby improving the prediction efficiency of protein residue motion trajectory coordinates. Attached Figure Description

[0036] 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.

[0037] 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.

[0038] Figure 1 This is a schematic diagram illustrating the principle of a traditional protein residue trajectory prediction scheme.

[0039] Figure 2 This is a flowchart illustrating an embodiment of the protein residue motion coordinate prediction method in this application.

[0040] Figure 3 This is a schematic diagram illustrating the principle of the protein residue motion coordinate prediction method in the embodiments of this application;

[0041] Figure 4This is a schematic diagram of the data flow of a feasible protein residue motion coordinate prediction scheme in the embodiments of this application;

[0042] Figure 5 This is a schematic diagram of the hardware operating environment involved in the protein residue motion coordinate prediction method in the embodiments of this application.

[0043] 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

[0044] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0045] 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.

[0046] 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.

[0047] In the field of protein dynamics prediction, this application proposes a novel approach for efficient modeling and inference of residue motion trajectories. Traditional protein residue motion trajectory prediction schemes, such as... Figure 1 As shown, the main approach is to add noise to the input protein structure based on a single-frame trajectory generation model, resulting in a noisy protein structure. Then, a denoising network (ScoreNetwork) is used to reconstruct the protein structure frame by frame, and this process is repeated N times frame by frame to obtain the reconstructed trajectory. It is evident that traditional prediction schemes often predict the 3D coordinates of each time point frame by frame in the time domain. As the number of prediction steps increases, the dimensions of the model's input and output increase linearly, leading to a surge in computation and a significant extension of inference time, making it difficult to meet the requirements for long-term conformational changes. The core focus of this application's embodiments is to extract the low-frequency components of protein residues over time from a frequency domain perspective through Fourier transform. These low-frequency components can reflect the main trends in the overall protein conformation, while high-frequency details contribute relatively little to macroscopic motion. Therefore, the technical solution of this application transforms the modeling object of the residue trajectory from hundreds to thousands of coordinate points in the original space into a small number of low-frequency coefficients that only need to be predicted (e.g., denoted as K-dimensional, much smaller than the time step T), thereby fundamentally reducing the dimensions that the model needs to process and reducing the computational load.

[0048] This application provides a method for predicting the motion coordinates of protein residues, referring to... Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the protein residue motion coordinate prediction method of this application. The protein residue motion coordinate prediction method includes:

[0049] Step S10: Obtain the first spectrum corresponding to the coordinate sequence of protein residues, and obtain the feature components within a preset frequency range from the first spectrum to determine the frequency domain feature vector, wherein the preset frequency range is lower than a preset frequency threshold.

[0050] In this context, the coordinate sequence of protein residues refers to the three-dimensional coordinate sequence of protein residues changing over time, while the first spectrum reflects the frequency domain characteristics of the coordinate sequence. The first spectrum can be obtained by performing a time-frequency domain transformation on the three-dimensional coordinate sequence. In this embodiment, the frequency domain feature vector in the first spectrum is used to replace the original coordinate sequence of protein residues for motion trajectory coordinate prediction. This reduces the originally high-dimensional problem that grows linearly with time steps to a fixed K-dimensional space, effectively reducing the computational load.

[0051] For example, all structural coordinates need to be centered and normalized before input to eliminate the impact of differences in protein size and coordinate scale on model training and prediction.

[0052] Furthermore, the feature classification within the preset frequency range in the first spectrum refers to the low-frequency components in the first spectrum, whose frequencies are greater than a preset frequency threshold. For example, for the first spectrum, several of the lowest frequency feature components can reflect the macroscopic motion of protein residues, while other feature components only reflect microscopic motion features to a certain extent, such as minute oscillations or noise, which can be ignored. In this embodiment, only the low-frequency components in the first spectrum are extracted to obtain the frequency feature vector, which compresses the dimensionality of the motion trajectory features of protein residues and also reduces the scale of the model's input and output data, thereby reducing the computational overhead of model training and inference.

[0053] This application's embodiments mainly overcome the limitations of high-dimensional time-domain modeling by using a dimensionality reduction approach. Specifically, this application's embodiments transform the spatial coordinate sequence of protein residues over time in the frequency domain, selecting only low-frequency feature components containing major conformational information for learning. This allows the modeling dimension to be reduced from T to a dimension K much smaller than T while maintaining prediction accuracy, thereby significantly reducing the scale of model input and output and thus significantly reducing the computational overhead of training and inference.

[0054] Step S20: Input the structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues into the preset potential diffusion model. The potential diffusion model performs diffusion prediction on the protein residues and determines the predicted frequency domain vector.

[0055] This application further incorporates the Latent Diffusion Model (LDM) technique, a technology that has demonstrated powerful capabilities in high-dimensional generation tasks in recent years. Unlike direct point-by-point generation based on the original coordinate sequence of protein residues, LDM can iteratively add and denoise low-frequency coefficients (i.e., frequency domain feature vectors) to generate a complete frequency domain representation (i.e., predicted frequency domain vector) in one step. By completing the diffusion process in the latent space, the model can effectively capture the global laws of residue movement while avoiding redundant calculations for independent inference in each frame, significantly shortening the overall inference time.

[0056] Furthermore, due to the complex spatial relationships inherent in the three-dimensional arrangement of protein residues, the diffusion generation process requires the use of Invariant Point Attention (IPA) to incorporate the spatial priors of the initial structure (specifically, structural feature vectors and sequence feature vectors of the encoded protein residues) into the model. This ensures that the generated predicted frequency domain vectors maintain reasonable conformational coherence when reconstructed into three-dimensional coordinates. This allows the model to significantly reduce dimensionality and accelerate inference while still outputting physically consistent and biologically interpretable dynamic trajectories. Therefore, the model simultaneously references the initial geometric priors in each denoising iteration to ensure that the generated low-frequency coefficients conform to both frequency domain trends and spatial structural consistency. The diffusion process completes the random noise injection and denoising restoration of all K low-frequency coefficients in the latent space in a single step, without requiring frame-by-frame iteration.

[0057] Step S30: Convert the predicted frequency domain vector into the corresponding second spectrum, and restore the second spectrum to the predicted coordinate sequence of protein residues.

[0058] After the potential diffusion model predicts the low-frequency frequency domain vector, it needs to be restored to obtain the motion trajectory in the form of a coordinate sequence. Specifically, referring to the aforementioned process of coordinate sequence transformation, the predicted frequency domain vector can be converted into the corresponding second spectrum, and then further restored into the predicted coordinate sequence of protein residues. The predicted coordinate sequence can intuitively show the motion trajectory of the protein residues.

[0059] The technical solution of this application maps the trajectory of protein residues to the frequency domain, retains only a few low-frequency frequency domain feature vectors, and generates a complete trajectory in the latent space all at once, thereby significantly reducing the model dimensionality and the number of iterations. Compared with the traditional time-domain frame-by-frame prediction scheme, the inference efficiency is improved by several times, and the time cost increases gradually with the trajectory length without sudden increases, achieving efficient expansion. In addition, this application embodiment also continuously integrates three-dimensional spatial constraints into the diffusion model with IPA attention, ensuring that the trajectory obtained after reconstruction of the generated prediction neighborhood vector can not only truly reflect the global motion trend of the protein, but also avoid structural drift and numerical instability.

[0060] Further, the step of obtaining the first spectrum corresponding to the coordinate sequence of protein residues, and obtaining the feature components within a preset frequency range from the first spectrum to determine the frequency domain feature vector may include:

[0061] Step S11: Perform a fast Fourier transform on the coordinate sequence of protein residues to obtain the first spectrum of the time-series signal corresponding to each of the three dimensions.

[0062] Step S12: Extract a preset number of low-frequency components with the largest amplitude from each first spectrum, or select low-frequency components within a preset frequency range from each first spectrum.

[0063] Step S13: Based on the protein residues and the order of the three-dimensional coordinate axes, the low-frequency components are spliced ​​into a frequency domain feature vector.

[0064] In this embodiment, a Fast Fourier Transform (FFT) is first performed on each residue coordinate sequence to calculate the first spectrum corresponding to the three time-series signals in the three-dimensional directions x, y, and z. The K low-frequency components with the largest amplitudes (or low-frequency components within a uniformly selected frequency range) are then extracted and concatenated according to the residue and coordinate axis order to a length of 3·K·N. e The frequency domain eigenvectors, where N e The number of residues is used. This frequency domain feature vector is used as the target representation of the potential diffusion model, significantly compressing the original T·3·N compared to the unprocessed 3D coordinate sequence. e Dimensional time-series signals.

[0065] Through the above steps, this embodiment of the application, while maintaining the accuracy of protein dynamic prediction, increases the modeling dimension from T·3·N. e Compressed to 3 kN e This significantly reduces the computational load and enables one-time complete trajectory generation with the help of a potential diffusion model, thus keeping the inference time within an acceptable range even for trajectory prediction tasks at the hundred-step level, thereby improving processing efficiency.

[0066] In this embodiment, the three-dimensional coordinate sequence of protein residues over time is first mapped to the frequency domain through discrete Fourier transform, retaining only the low-frequency coefficients containing macroscopic motion information, thereby reducing the high-dimensional problem that originally grew linearly with time steps to a fixed K-dimensional space.

[0067] In one feasible embodiment, the step of converting the predicted frequency domain vector into a corresponding second spectrum and restoring the second spectrum to a predicted coordinate sequence of protein residues may include:

[0068] Step S31: Perform high-frequency zero-filling on the predicted frequency domain vector to obtain the second spectrum;

[0069] Step S32: Perform an inverse fast Fourier transform on the second spectrum to obtain the predicted coordinate sequence of protein residues.

[0070] Zero-padding the predicted frequency domain vector in the frequency domain to achieve spectral completion is primarily based on the characteristics of the Fourier transform in signal processing and its time-frequency domain correspondence. Mathematically, it involves inserting zero values ​​in the middle (high-frequency band) of the frequency domain signal, effectively expanding the original spectrum. Due to the symmetry of the Fourier transform, zero-padding in the frequency domain does not change the actual frequency components of the signal, but it generates a higher number of sampling points in the time domain through the inverse Fast Fourier transform (IFFT). For example, if the original frequency domain signal length is N, after zero-padding it becomes M (M>N), then the number of sampling points in the time domain also becomes M, equivalent to interpolating the original time domain signal.

[0071] After restoring the reconstructed low-frequency coefficients to the complete spectrum, the frequency domain data is then restored to a T-step coordinate sequence in time domain form through inverse fast Fourier transform, which intuitively reflects the motion trajectory information of protein residues.

[0072] In conjunction with the foregoing embodiments, this application essentially provides a frequency domain shallow space trajectory generation model, such as... Figure 3 As shown, the input protein structure is first subjected to a Fourier transform to obtain K-dimensional low-frequency coefficients (noise), with a size of ×2×K. Then, the structure feature vector and sequence feature vector encoded by the encoder are input into the latent space diffusion model (LDM) along with the aforementioned low-frequency coefficients (equivalent to frequency domain feature vectors). The decoder then decodes the LDM and outputs the predicted K-dimensional low-frequency coefficients, with a size of ×2×K. Finally, the trajectory is reconstructed by inverse Fourier transform, thus realizing the trajectory reconstruction of the input protein structure based on low-frequency dynamics.

[0073] In one feasible embodiment, such as Figure 4As shown, the potential diffusion model includes a normalization module and a flow network module;

[0074] The step of inputting the structural feature vector, sequence feature vector, and frequency domain feature vector of protein residues into a preset latent diffusion model, and having the latent diffusion model predict the diffusion of the protein residues to determine the predicted frequency domain vector includes:

[0075] Step S21: After the structural feature vector, sequence feature vector, and frequency domain feature vector of protein residues are input into the preset potential diffusion model, the frequency domain feature vector is standardized by the standardization module to obtain the standard frequency domain vector.

[0076] like Figure 4 As shown, the latent space of the latent diffusion model consists of two parts: a normalization module and a flow network module. The frequency domain feature components obtained from the motion spectrum are encoded by the encoder to obtain z0, and then normalized to obtain the standard frequency domain vector z. t Then the standard frequency domain vector z t The input is fed into a streaming network, which outputs a predicted frequency domain vector z. * The decoder (from 0 to Decoder) reconstructs the spectrum and then uses IFFT (Inverse Fast Fourier Transform) to reconstruct the trajectory and protein structure. The standardization process includes centering and normalization to eliminate the impact of differences in the size and coordinate scale of different proteins on model training and prediction.

[0077] Step S22: The standard frequency domain vector is processed by the residual block in the flow network module, and then the processed standard frequency domain vector, sequence feature vector and structural feature vector are fused into the attention calculation to obtain a multi-dimensional frequency domain feature vector.

[0078] Specifically, the core module based on the Latent Diffusion Model (LDM) combines a denoising network in the style of U-Net (a convolutional neural network architecture for image segmentation) with IPA. The network input of the streaming network module consists of three parts: the frequency domain features z after the current perturbation. tThe standard frequency domain vector consists of three parts: a standard frequency domain vector, a structural latent vector of protein residues generated by the encoder, and a sequence feature vector obtained by one-hot encoding or pre-training embedding of the amino acid sequence of the protein residues. The flow network first processes the standard frequency domain vector through several residual blocks. Then, in multiple IPA sublayers, it integrates the three-dimensional geometric information of the coordinate system (including the aforementioned three network inputs) into the attention calculation, ensuring that overall spatial consistency is preserved during denoising. Specifically, each diffusion time step t corresponds to a set of denoising parameters, with linear or cosine scheduling for noise control, and the denoising objective is the reconstruction estimation of the frequency domain feature vector.

[0079] Step S23: Initialize the multidimensional frequency domain feature vector based on the normal distribution, and perform denoising iteration on the multidimensional frequency domain feature vector to obtain the predicted feature vector.

[0080] During the inference phase of the flow network module, full-range diffusion sampling is performed on the K-dimensional frequency domain feature vector (i.e., multi-dimensional frequency domain feature vector) in the latent space. Specifically, the frequency domain coefficient noise is first initialized from a standard normal distribution, and after L-step denoising iterations, a clear frequency domain vector is obtained. It should be noted that this inference process does not require frame-by-frame iteration, but rather generates a complete frequency domain representation at once, obtaining the corresponding predicted feature vector that reflects the predicted trajectory of protein residues.

[0081] Furthermore, in a feasible embodiment, before the step of inputting the structural feature vector, sequence feature vector, and frequency domain feature vector of protein residues into a preset potential diffusion model, the method may further include:

[0082] Step A10: Obtain the initial static structure and amino acid sequence of the protein residues;

[0083] Step A20: The initial static structure is encoded using a preset variational autoencoder to obtain the structural feature vector of protein residues;

[0084] Step A30: Extract the sequence feature vectors corresponding to the amino acid sequences.

[0085] In this embodiment, the initial static structure of the protein residues (equivalent to step 0 structure) and the amino acid sequence are first obtained. The preset variational autoencoder can be a GearBind (a pre-trainable deep neural network) structure VAE (Variational Auto Encoder), which encodes the step 0 structure of the protein residues to obtain the corresponding structural feature vector. This VAE represents the protein structure as graph data; node features include residue or atom coordinates and chemical properties, while edge features represent the interactions between residues. The encoder part consists of a multi-layer graph convolutional network (GraphConv) and a variational layer, finally outputting a latent vector z, whose dimension is generally much smaller than the number of input nodes.

[0086] In another feasible embodiment, such as Figure 4 As shown, the entire frequency domain shallow space trajectory generation model also includes a decoder structure (i.e., a decoder), which is symmetrical to the encoder and is used for structural reconstruction and latent space regularization during training to reconstruct the spectrum. In the embodiments of this application, pre-training or fine-tuning with GearBind VAE allows the latent vectors to carry rich geometric priors.

[0087] In addition, when extracting the sequence feature vectors corresponding to amino acid sequences, the amino acid sequences can be encoded using the One-hot encoding method, or they can be processed using pre-trained Embedding to obtain the corresponding sequence feature vectors for input into the stream network module.

[0088] In this embodiment, by employing GearBind VAE to encode the initial static structure, high-quality geometric priors can be provided for LDM, resulting in compact geometric latent identifiers, which are then input into the latent diffusion model along with low-frequency feature vectors in the frequency domain. Simultaneously, an Invariant Point Attention (IPA) mechanism continuously introduces three-dimensional structural constraints during the diffusion denoising process, ensuring that the generated trajectory maintains physical plausibility and spatial consistency beyond the global motion trend. This approach balances prediction accuracy with maximizing the efficiency advantages of frequency domain dimensionality reduction and latent space generation. Therefore, by combining low-frequency modeling in the frequency domain with latent diffusion generation, the technical solution of this embodiment solves the problems of low efficiency and poor scalability in high-dimensional time-domain iteration of traditional protein dynamic prediction. It ensures prediction accuracy for long trajectories while keeping inference time within an acceptable range, thus meeting the demand for efficient simulation of large-scale, long-term resolution protein motion.

[0089] In one embodiment, the protein residue motion coordinate prediction method may further include:

[0090] Step B10: Obtain the degree features of protein residues and determine the degree features as the reconstruction content target of the initial diffusion model;

[0091] Step B20: Train the initial diffusion model based on the content reconstruction target to obtain a pre-trained model;

[0092] Step B30: Transfer the pre-trained model to the preset spectrum reconstruction task, and train the pre-trained model with the goal of minimizing the frequency domain reconstruction error to obtain the potential diffusion model.

[0093] This application provides a method for training a latent diffusion model. During model training, frequency domain reconstruction error (MSE Loss) is used as the primary objective, enabling the latent diffusion model to learn to accurately reconstruct low-frequency components and obtain a more accurate and stable predicted frequency domain vector.

[0094] Optionally, geometric consistency loss or skeleton preservation loss can be added to the training objective during training to further constrain the physical continuity between the denoised result and the initial structure. The entire model training is performed in parallel on a GPU (Graphics Processing Unit) cluster.

[0095] It should be noted that before pre-training, degree features are constructed as pre-training reconstruction content, and then the pre-trained model is transferred to the spectrum reconstruction task. The batch size during pre-training can be dynamically adjusted according to the protein size. The learning rate is adjusted using Adam (Adaptive Moment Estimation optimizer) in conjunction with cosine annealing and warm-up strategies.

[0096] In another embodiment, KL (Kullback-Leibler, relative entropy) divergence regularization needs to be applied to the pre-trained VAE (Variational Autoencoder) portion to make the latent space distribution tend towards a standard normal distribution, so as to conditionally diffuse the input. Figure 4 As shown, during the pre-training process, the graph features of the protein structure are mainly extracted and then encoded and decoded by an Incoder and a Decoder to complete the training process.

[0097] In one feasible embodiment, prior to the step of obtaining the first spectrum corresponding to the coordinate sequence of protein residues, the method may further include:

[0098] Step C10: Obtain the original three-dimensional structure data and motion trajectory data of the protein;

[0099] Step C20: Based on the motion trajectory data, record the three-dimensional coordinates of each protein residue in time series form to obtain the coordinate sequence of the protein residues.

[0100] When implementing the technical solutions of the embodiments of this application, such as Figure 4 As shown, the first step is to prepare the data required for training and inference, including the original three-dimensional structural coordinates of the protein and the corresponding motion trajectory data (i.e., the dataset trajectory). The protein sequence includes sequences such as "HE…GTQ". Before formally predicting the motion coordinates of protein residues, the complete T-step trajectory is obtained from molecular dynamics simulations or high-quality experimental data, and the x, y, z, and other three-dimensional coordinates of each residue are recorded in time series form. During the model inference stage, only the static structure and amino acid sequence of step 0 are required; step 0 refers to the initial state.

[0101] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the protein residue motion coordinate prediction method of this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0102] This application also provides a protein residue motion coordinate prediction device, which includes at least: 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, which are executed by the at least one processor to enable the at least one processor to perform the protein residue motion coordinate prediction method in the above embodiments.

[0103] The following is for reference. Figure 5 It shows a schematic diagram of the structure of a control unit suitable for implementing the embodiments of this application. Figure 5 The control unit shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this application.

[0104] like Figure 5As shown, the control unit may include a processing device 101 (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 read-only memory (ROM) 102 or a program loaded from storage device 103 into random access memory (RAM) 104. RAM 104 also stores various programs and data required for the operation of the control unit. The processing device 101, ROM 102, and RAM 104 are interconnected via bus 105. Input / output (I / O) interface 106 is also connected to the bus. Typically, the following systems can be connected to I / O interface 106: input devices 107 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 108 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 103 including, for example, magnetic tapes, hard disks, etc.; and communication devices 109. Communication device 109 allows the control unit to communicate wirelessly or wiredly with other devices to exchange data. Although the diagram shows control units 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.

[0105] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure 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 103, or installed from ROM 102. When the computer program is executed by processing device 101, it performs the functions defined in the methods of the embodiments of this application.

[0106] The protein residue motion coordinate prediction device provided in this application, employing the protein residue motion coordinate prediction method described in the above embodiments, can solve the technical problems of high computational load and low efficiency in traditional protein residue motion trajectory prediction schemes. Compared with the prior art, the beneficial effects of the protein residue motion coordinate prediction device provided in this application are the same as those of the protein residue motion coordinate prediction method provided in the above embodiments, and other technical features in this protein residue motion coordinate prediction device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0107] It should be understood that various parts of the embodiments of 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.

[0108] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments 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 protection scope of the embodiments of this application. Therefore, the protection scope of the embodiments of this application should be determined by the protection scope of the above claims.

[0109] This application also provides a computer-readable storage medium storing a computer program that can run on a processor. The computer program is used to execute the protein residue motion coordinate prediction method in the above embodiments.

[0110] The computer-readable storage medium provided in this application embodiment may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or 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 fibers, 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.

[0111] The aforementioned computer-readable storage medium may be included in the protein residue motion coordinate prediction device; or it may exist independently and not assembled into the protein residue motion coordinate prediction device.

[0112] The aforementioned computer-readable storage medium carries one or more programs. When the one or more programs are executed by the protein residue motion coordinate prediction device, the protein residue motion coordinate prediction device performs the following: acquiring a first spectrum corresponding to the coordinate sequence of protein residues, and acquiring feature components within a preset frequency range from the first spectrum to determine a frequency domain feature vector, wherein the preset frequency range is lower than a preset frequency threshold; inputting the structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues into a preset latent diffusion model, and having the latent diffusion model perform diffusion prediction on the protein residues to determine a predicted frequency domain vector; converting the predicted frequency domain vector into a corresponding second spectrum, and restoring the second spectrum to the predicted coordinate sequence of the protein residues.

[0113] Computer program code for performing the operations of this disclosure 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).

[0114] 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, may 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.

[0115] 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.

[0116] The computer-readable storage medium provided in this application embodiment stores computer-readable program instructions for executing the above-described protein residue motion coordinate prediction method, which can solve the technical problems of large computational load and low efficiency in traditional protein residue motion trajectory prediction schemes. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application embodiment are the same as the beneficial effects of the protein residue motion coordinate prediction method provided in the above embodiments, and will not be repeated here.

[0117] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the protein residue motion coordinate prediction method described above.

[0118] The computer program product provided in this application can solve the technical problems of high computational load and low efficiency in traditional protein residue motion trajectory prediction schemes. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the protein residue motion coordinate prediction method provided in the above embodiments, and will not be repeated here.

[0119] The above are merely exemplary solutions of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent processing scope of this application.

Claims

1. A method for predicting the motion coordinates of protein residues, characterized in that, The method for predicting the motion coordinates of protein residues includes: Obtain the first spectrum corresponding to the coordinate sequence of protein residues, and extract the feature components within a preset frequency range from the first spectrum to determine the frequency domain feature vector, wherein the preset frequency range is lower than a preset frequency threshold; The structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues are input into a preset potential diffusion model, and the potential diffusion model performs diffusion prediction on the protein residues to determine the predicted frequency domain vector. The predicted frequency domain vector is converted into the corresponding second spectrum, and the second spectrum is restored to the predicted coordinate sequence of protein residues; The step of obtaining the first spectrum corresponding to the coordinate sequence of protein residues, and obtaining feature components within a preset frequency range from the first spectrum to determine the frequency domain feature vector includes: A fast Fourier transform is performed on the coordinate sequence of the protein residues to obtain the first spectrum of the time-series signal corresponding to each of the three dimensions; Extract a preset number of low-frequency components with the largest amplitude from each of the first spectra, or select low-frequency components within a preset frequency range from each of the first spectra; The low-frequency components are concatenated into a frequency domain feature vector based on the protein residues and the three-dimensional coordinate axis order.

2. The protein residue motion coordinate prediction method as described in claim 1, characterized in that, The potential diffusion model includes a normalization module and a flow network module; The step of inputting the structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues into a preset latent diffusion model, and having the latent diffusion model predict the diffusion of the protein residues to determine the predicted frequency domain vector includes: After the structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues are input into a preset potential diffusion model, the frequency domain feature vector is standardized by the standardization module to obtain a standard frequency domain vector. The standard frequency domain vector is processed by the residual block in the flow network module, and then the processed standard frequency domain vector, the sequence feature vector, and the structural feature vector are fused into the attention calculation to obtain a multi-dimensional frequency domain feature vector. The multidimensional frequency domain feature vector is initialized based on a normal distribution, and the multidimensional frequency domain feature vector is denoised iteratively to obtain the predicted feature vector.

3. The protein residue motion coordinate prediction method as described in claim 2, characterized in that, Before the step of inputting the structural feature vector, sequence feature vector, and frequency domain feature vector of the protein residues into a preset potential diffusion model, the method further includes: To obtain the initial static structure and amino acid sequence of protein residues; The initial static structure is encoded by a preset variational autoencoder to obtain the structural feature vector of the protein residues; Extract the sequence feature vector corresponding to the amino acid sequence.

4. The protein residue motion coordinate prediction method as described in claim 2, characterized in that, The method for predicting the motion coordinates of protein residues also includes: The degree features of protein residues are obtained, and these degree features are determined as the reconstruction content target of the initial diffusion model. The initial diffusion model is trained based on the reconstructed content target to obtain a pre-trained model; The pre-trained model is transferred to a preset spectrum reconstruction task, and the pre-trained model is trained with the goal of minimizing the frequency domain reconstruction error to obtain a potential diffusion model.

5. The protein residue motion coordinate prediction method as described in claim 1, characterized in that, The step of converting the predicted frequency domain vector into a corresponding second spectrum and restoring the second spectrum into a predicted coordinate sequence of protein residues includes: The predicted frequency domain vector is filled with high-frequency zeros to obtain the second spectrum; Perform an inverse fast Fourier transform on the second spectrum to obtain the predicted coordinate sequence of protein residues.

6. The protein residue motion coordinate prediction method according to any one of claims 1 to 5, characterized in that, Prior to the step of obtaining the first spectrum corresponding to the coordinate sequence of protein residues, the method further includes: Acquire raw three-dimensional structural data and motion trajectory data of proteins; Based on the motion trajectory data, the three-dimensional coordinates of each protein residue are recorded in time series form to obtain the coordinate sequence of the protein residues.

7. A protein residue motion coordinate prediction device, characterized in that, The protein residue motion coordinate prediction device includes at least: 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 protein residue motion coordinate prediction method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for implementing a method for predicting the motion coordinates of protein residues, the program for implementing the method for predicting the motion coordinates of protein residues being executed by a processor to implement the steps of the method for predicting the motion coordinates of protein residues as described in any one of claims 1 to 6.

9. 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 protein residue motion coordinate prediction method as described in any one of claims 1 to 6.

Images