Protein conformation change prediction method based on deep generative model and reinforcement learning

Abstract

The invention relates to a protein conformation change prediction method based on a deep generative model and reinforcement learning. The method comprises the following steps: acquiring a plurality of sample data to form a sample data set; masking the (K + 1)-th frame to the L-th frame of molecular dynamics simulation trajectory data of each piece of sample data in the sample data set to obtain a masked sample data set; inputting the sample data set, the masked sample data set and the first condition guide information into a to-be-trained protein conformation prediction model for training, and performing fine tuning on network parameters of the trained protein conformation prediction model by adopting a reinforcement learning fine tuning algorithm optimized by a protein conformation dynamics strategy, and obtaining a trained protein conformation prediction model. And obtaining L-frame random noise, first condition guidance information and K-frame protein initial conformations, and inputting the L-frame random noise, the first condition guidance information and the K-frame protein initial conformations into the trained protein conformation prediction model for prediction to obtain L-frame protein conformation change trajectories. Therefore, the protein conformation change process can be dynamically analyzed.

Description

Technical Field

[0001] This application relates to the fields of computational biology and artificial intelligence, and in particular to a method for predicting protein conformational changes based on deep generative models and reinforcement learning. Background Technology

[0002] In molecular biology, proteins, as the ultimate executors of life functions, have functional specificity determined by their three-dimensional spatial structure. The specific three-dimensional arrangement formed by their folding, i.e., protein conformation, is the structural basis for their biological activity. However, modern biophysics suggests that proteins are not rigid entities under physiological conditions, but rather follow the principles of statistical mechanics, constantly undergoing thermodynamic fluctuations in their energy landscape. This means that every protein exists as a conformational ensemble, a dynamic set containing countless instantaneous conformational states. Protein functions, such as allosteric regulation, enzyme catalysis, and induced fit with ligands, are all achieved through state transitions within this conformational ensemble. Therefore, predicting the conformational changes of proteins over long timescales, or accurately characterizing protein conformational dynamics, is of crucial scientific significance and practical value for fundamentally elucidating molecular mechanisms and discovering "hidden" drug targets exposed only in rare conformations.

[0003] The current existing technologies related to protein conformation are as follows:

[0004] (1) First-principles-driven molecular dynamics (MD) simulations are the theoretical cornerstone of protein dynamics research. They numerically solve Newton's equations of motion based on classical force fields, tracking the evolution of the system over time at atomic resolution, theoretically providing the most detailed dynamic information. Problems or limitations: 1) Time scale limitation: The time step of MD simulations is limited to femtoseconds (10^6) seconds. -15 s) or picosecond (10 -12 The timescale is on the order of s, while most key biological processes occur in microseconds (10^6 seconds). -61) On longer timescales, the huge gap leads to computationally prohibitive direct predictive modeling of these processes, constituting the so-called "timescale problem." Although some downstream tasks only require a set of dynamic conformations with a large degree of coarseness within a certain stage of protein conformational change, MD methods often require full timescale modeling based on femtosecond or picosecond scales. This also results in a significant waste of computational resources in MD simulations of large-scale proteins; 2) High professional threshold and parameter sensitivity: MD methods are not "plug and play." They often require users to have deep professional knowledge to select appropriate collective variables or set complex parameters (such as simulating force fields, various conditions that need to be set to simulate physiological states). The quality of parameter selection directly determines the success or failure of the simulation, which greatly reduces the versatility and automation of these methods; 3) The computational cost remains high: Although some augmented sampling methods can shorten the simulation time to some extent by overcoming energy barriers, the computational cost and resource consumption are still unacceptable for ordinary researchers who lack high-performance GPU computing clusters.

[0005] (2) Extension methods based on deep learning and technically extending static structure prediction models, such as AlphaFold2, are based on predicting the single most energy-favorable static conformation of an amino acid sequence. In recent years, researchers have also explored using byproduct information from these models (such as the confidence level of the Predicted Local Distance Difference Test (PLDDT)) or introducing random search and flexible region reasoning into the prediction. Studies have shown that even deep learning-based static structure prediction methods like AlphaFold have the potential to reveal and generate multiple different conformations from a single protein sequence by capturing evolutionary and structural features. Problems: These methods are essentially post-processing or extensions of static prediction models, and their inherent framework determines their fundamental limitations: the set of conformations they generate is a collection of temporally unrelated, discrete "snapshots". These methods cannot adequately search and sample the conformational dynamics space, nor can they model the physical processes of conformational change itself. Therefore, they cannot provide any information about the pathways, kinetic rates, energy barriers, or temporal dependencies between states during conformational transitions. In short, they can predict "what conformations a protein might have," but they cannot model the core dynamic question of "how a protein transitions from one conformation to another."

[0006] (3) The method based on deep generative models and searching and sampling the protein conformational dynamics space aims to learn the implicit conformational dynamics space distribution within protein molecular dynamics (MD) simulation data (i.e., learn its Boltzmann distribution in energy topography), and based on this, quickly sample and generate new conformations that conform to this distribution. Problems or shortcomings: This type of generative model has a fundamental limitation, resulting in a lack of ability to dynamically analyze and model protein conformational changes: that is, they focus on learning static probability distributions. The core goal of these models is to learn and reproduce the static probability distribution of a conformational space, rather than the dynamic process of conformational evolution over time. At the same time, they lack effective guidance mechanisms oriented towards process prediction. Such models are often based on unconditional generation methods or weakly guided generation based solely on the static information of protein residue sequences. Therefore, these generative models, based on relatively weak or even uncontrollable guidance mechanisms and inherently lacking the ability to learn dynamic behavior, cannot meet the need to predict protein conformational changes, constituting a technological gap that urgently needs to be addressed in the current field.

[0007] Therefore, current models lack the ability to dynamically analyze protein conformational changes: they focus on learning static probability distributions, with the core objective being to learn and reproduce the static probability distribution of a conformational space, rather than the dynamic process of conformational evolution over time. Furthermore, they lack effective guidance mechanisms for process prediction. Existing models often rely on unconditional generation methods or weakly guided generation based solely on the static information of protein residue sequences. Thus, these generative models, based on relatively weak or even uncontrollable guidance mechanisms and inherently lacking the ability to learn dynamic behavior, cannot meet the need for predicting protein conformational changes, constituting a pressing technological gap in the field. Summary of the Invention

[0008] Therefore, it is necessary to provide a protein conformation change prediction method based on deep generative models and reinforcement learning that can dynamically analyze the protein conformation change process, addressing the aforementioned technical problems.

[0009] A method for predicting protein conformational changes based on deep generative models and reinforcement learning, the method comprising:

[0010] The L-frame random noise, first condition guidance information, and second condition guidance information are obtained. The first condition guidance information includes protein residue sequence, protein spatial contact map, and protein evolution specificity scoring matrix. The second condition guidance information is the K-frame protein initial conformation.

[0011] The L-frame random noise, the first conditional guidance information, and the K-frame initial protein conformation are input into the trained protein conformation prediction model for prediction to obtain the L-frame protein conformation change trajectory.

[0012] The training method for the protein conformation prediction model is as follows:

[0013] Molecular dynamics simulation trajectory data of different proteins with fewer than 77 residues within 10 ns under physiological conditions and molecular dynamics simulation trajectory data of different proteins with more than 77 residues to 200 residues were obtained and preprocessed to obtain multiple sample data to form a sample dataset, each sample data including L frames.

[0014] The molecular dynamics simulation trajectory data of each sample data in the sample dataset from frame K+1 to frame L are masked to obtain the masked sample dataset, which includes the masked sample data corresponding to each sample data.

[0015] The sample dataset, the masked sample dataset, and the first condition guidance information are input into the protein conformation prediction model to be trained to obtain the trained protein conformation prediction model.

[0016] L frames of Gaussian noise are randomly sampled, and the L frames of Gaussian noise, the first conditional guidance information and the third conditional guidance information are input into the trained protein conformation prediction model to predict the conformational change trajectory, and obtain the protein conformational change trajectory prediction result.

[0017] Based on the predicted protein conformational change trajectory, a reinforcement learning fine-tuning algorithm optimized by a protein conformational dynamics strategy is used to fine-tune the network parameters of the trained protein conformational prediction model, thereby obtaining a well-trained protein conformational prediction model.

[0018] Compared with the prior art, this application achieves the following significant beneficial effects through the above technical solution:

[0019] 1. Achieved end-to-end prediction of complete dynamic processes, filling a technological gap: This application is the first protein conformation prediction model that can directly predict and generate a complete, temporally coherent protein conformational change trajectory from a given initial protein conformation. It successfully transforms the problem from "predicting possible states" to "predicting complete dynamic processes," solving the core pain point that existing static models cannot provide dynamic path information.

[0020] 2. Significantly improves computational efficiency and overcomes the "time scale" bottleneck: Compared to traditional MD simulations that require massive atomic-level integration operations, the protein conformation prediction model in this application, as a deep learning proxy model, has an extremely fast inference speed once trained. It can generate dynamic information at the microsecond level or higher, which usually requires days or even months of MD simulations, within minutes or hours, greatly reducing computational and time costs.

[0021] 3. Significantly improved physical realism of prediction results through reinforcement learning: A key breakthrough is the original reinforcement learning fine-tuning algorithm for optimizing protein conformation dynamics strategies. It optimizes the protein conformation prediction model online through a meticulously designed comprehensive reward function incorporating multiple physical and biological constraints. This results in a final protein conformation ensemble that better conforms to real physicochemical laws in terms of energy, interatomic spacing, and secondary structure compared to existing generative models, thus solving the problem of generative models easily producing unreasonable structures.

[0022] 4. Conditioned and controllable predictive generation is achieved, enhancing practical value: The conditional generation mechanism of this application makes it highly practical. Researchers can specify the initial set of protein conformations of interest and use this application to analyze and predict their subsequent dynamic changes, which has important guiding significance for understanding specific biological issues such as drug action mechanisms and allosteric regulation.

[0023] 5. Lowered barrier to entry and improved the method's versatility and automation: As a pre-trained end-to-end model, this application is far simpler to use than complex MD simulations or augmented sampling methods. Users do not need complex parameter settings or a deep theoretical background to perform kinetic predictions of target proteins, greatly enhancing the technology's popularity and application potential among a wide range of biomedical researchers.

[0024] In summary, this application provides a novel, efficient, and physically realistic computational paradigm for protein conformation dynamics research through the innovative combination of deep generative models and reinforcement learning, effectively overcoming multiple limitations of existing technologies. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a protein conformation change prediction method based on deep generative models and reinforcement learning in one embodiment.

[0026] Figure 2 This is a schematic diagram of the network structure of the noise prediction module in one embodiment;

[0027] Figure 3 This is an Atom14 atomic representation standard for grouping and filling key atoms within a protein by residue in one embodiment.

[0028] Figure 4 This is a schematic diagram of the protein backbone twist angle in one embodiment;

[0029] Figure 5 This is a representation of the seven-element torsion angle of protein residues in one embodiment;

[0030] Figure 6 This is a schematic diagram of the forward noise addition and backward noise reduction process of a protein conformation prediction model in one embodiment.

[0031] Figure 7 This is a schematic diagram of the denoising process at each diffusion time step in one embodiment. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0033] In one embodiment, such as Figure 1 As shown, a method for predicting protein conformational changes based on deep generative models and reinforcement learning is provided, including the following steps:

[0034] Step S220: Obtain L-frame random noise, first condition guidance information, and second condition guidance information. The first condition guidance information includes protein residue sequence, protein spatial contact map, and protein evolution specificity scoring matrix. The second condition guidance information is the K-frame protein initial conformation.

[0035] The protein conformation prediction model cannot directly input protein conformations represented in Cartesian coordinates; instead, the protein conformation must be converted to a torsion angle representation. Therefore, the initial protein conformation is a protein conformation that has been converted to a torsion angle representation.

[0036] The conversion process from protein conformation to torsion angle representation is as follows:

[0037]

[0038] Where, pre-ω, ψ, χ1, χ2, χ3, χ4 represent seven types of torsion angles. The first three are the torsion angles formed by the Cα of a residue (not at the beginning or end) on the main chain with the N atom of the preceding peptide bond, the C atom of the subsequent peptide bond, and the Cα of the subsequent residue. The remaining four are the torsion angles that the side chain may form (the side chain state of each amino acid residue may be different). Let any twist angle be represented by a tuple after the transformation; It represents the set of real numbers.

[0039] Finally, for each torsion angle, it can be represented based on sine and cosine encoding:

[0040]

[0041] Here, θ represents the set of twist angles for the entire protein, which is represented by tuples of twist angles for the corresponding number of residues. The set is mainly represented in matrix form; This represents any torsion angle after sine and cosine encoding. That is, sine and cosine encoding is performed on any torsion angle tuple representation in θ to obtain the corresponding encoded torsion angle representation.

[0042] Step S240: Input the random noise of frame L, the first condition guidance information, and the initial protein conformation of frame K into the trained protein conformation prediction model to predict and obtain the protein conformation change trajectory of frame L.

[0043] It should be understood that the initial conformation of the protein in the K frames of the well-trained protein conformation prediction model is represented by the torsion angle, and the corresponding conformational change trajectory of the protein in the L frames is also represented by the torsion angle. After obtaining the conformational change trajectory of the protein in the L frames, coordinate inverse mapping should be performed to visualize the conformational change trajectory of the protein.

[0044] The inverse coordinate mapping includes:

[0045] (1) Input data definition: The protein conformational change trajectory is represented by the torsion angle as follows: Where L is the number of frames in the protein conformational change trajectory; E is the number of residues in the current protein; and 7 is the number of torsion angle types. 2 represents the sin / cos encoding dimension;

[0046] The formula for decoding the torsion angle is:

[0047] μ i,j =atan2(sinμ i,j ,cosμ i,j )

[0048] Where, μ i,j Let be the j-th torsion angle of the i-th residue.

[0049] (2) Coordinate reconstruction based on the Natural Extension Reference Frame algorithm: The Natural Extension Reference Frame algorithm is used for 3D coordinate reconstruction, and the inverse-transformed atomic coordinates are organized into the standard Atom14 format.

[0050]

[0051] Catom14 This is the standard for representing the key heavy atom group of Atom14 residues, c 1,1:14 The three-dimensional coordinate representation of the L-frames representing the 1st to 14th critical heavy atoms within the first residue of a protein, c 2,1:14 The three-dimensional coordinate representation of the L-frame representing the first to fourteenth critical heavy atoms within the second residue of a protein, c E,1:14 The three-dimensional coordinate representation of the L-frame representing the first to the fourteenth critical heavy atoms within the E-th residue of a protein.

[0052] (3) Mapping of all-atomic Cartesian coordinates

[0053] Based on the inverse transformation of Atom14 coordinates, the Pyrosetta molecular modeling software library was used to refine and optimize the all-atomic structure. This process includes the following steps: Pose object initialization, skeleton coordinate mapping, side-chain optimization, and hydrogen atom addition.

[0054] The trained protein conformation prediction model can be constructed through a phased "training phase-fine-tuning phase" framework. This model is an end-to-end deep learning model that can use the initial conformation of K frames of proteins and the first conditional guidance information as conditional guidance, and performs better in terms of physical properties and temporal coherence.

[0055] It should be understood that, guided by the initial conformation of proteins in K frames and the first condition guidance information, a well-trained protein conformation prediction model can predict end-to-end long-scale protein conformation change trajectories that exhibit better physical properties and temporal coherence.

[0056] In one embodiment, the protein conformation prediction model includes a multimodal local spatial perception module, a noise prediction module, and a denoising module.

[0057] To provide rich biological prior knowledge for the protein conformation prediction model, this application proposes a Multi-Model Local Aware-Spatial Transformer (MLAST) network architecture. Before analysis and prediction begin, this architecture integrates multi-modal information from large protein language models (such as ESM2), including protein sequence features, protein spatial contact maps, and protein evolution-specific scoring matrices, generating a highly information-condensed, spatially aligned sequence feature. It also enables persistent cross-transfer and fusion of information with the noise prediction module.

[0058] The noise prediction module is a persistent spatiotemporal Transformer network used to predict noise.

[0059] The noise prediction module consists of multiple persistent spatiotemporal attention layers stacked together. Each layer alternately performs attention calculations in both the spatial dimension (between residues) and the temporal dimension (between frames). In particular, during spatial attention calculations, enhanced sequence features (i.e., fused protein dynamic features) provided by the multimodal local spatial awareness module are cross-integrated again. This design, known as persistent spatial awareness attention, ensures that biophysical constraints continuously guide the entire denoising process.

[0060] The protein conformation prediction model generates a complete trajectory in one go using a non-autoregressive approach. To ensure the temporal consistency of the trajectory and its dependence on the initial state, a conditional mask-guided training mechanism is designed. During training, the first frame (or the first few frames) of a complete molecular dynamics simulation trajectory data is used as a condition, and these frames are marked as known information using a mask. The task of the protein conformation prediction model then becomes: given these known initial frames, predict and reconstruct the entire molecular dynamics simulation trajectory data "contaminated" by noise. This forces the protein conformation prediction model to learn the ability to perform dynamic deductions starting from a specific starting point.

[0061] After training, the protein conformation prediction model adopts a deterministic backsampling strategy. Starting from random noise and a given initial protein conformation, it efficiently generates a complete and deterministic protein conformation change trajectory through iterative denoising.

[0062] In one embodiment, L-frame random noise, first conditional guidance information, and K-frame initial protein conformation are input into a trained protein conformation prediction model to predict the L-frame protein conformation change trajectory, including:

[0063] Protein residue sequences, protein spatial contact maps, and protein evolution-specific scoring matrices are input into a multimodal local spatial perception module for feature fusion and enhancement to obtain fused protein dynamic features. The fused protein dynamic features, L-frame random noise, and K-frame initial protein conformation are input into a noise prediction module, which obtains noise prediction results under the guidance of the fused protein dynamic features and K-frame initial protein conformation. Based on the noise prediction results, the L-frame random noise is denoised to obtain the L-frame protein conformation change trajectory.

[0064] The noise prediction module is a highly efficient diffusion-generative model based on autoencoder mask training and inference. Its fundamental task is to receive a complete protein trajectory "damaged" by noise and learn how to "recover" the original, "clean" trajectory (specifically, by predicting the added noise). This recovery process is strongly guided by the undamaged conformational frames used as initial conditions. As a unified whole, it simultaneously performs feature extraction (encoding) of the input information and reconstruction (decoding) of the target trajectory.

[0065] The noise prediction module is a spatiotemporal Transformer-based noise prediction network that performs cross-self-attention calculations at both the protein main chain structure and time series scales, thereby achieving better adaptive noise distribution prediction during the denoising process.

[0066] In one embodiment, the protein residue sequence, protein spatial contact map, and protein evolution-specific scoring matrix are input into a multimodal local spatial awareness module for feature fusion and enhancement to obtain fused protein dynamic features, including:

[0067] The protein residue sequence is input into the large prediction model of the multimodal local spatial perception module for feature extraction to obtain protein sequence features. The protein sequence features and protein spatial contact map are fused by the feature fusion unit of the multimodal local spatial perception module, and then self-attention is calculated by the self-attention unit of the multimodal local spatial perception module to obtain spatially aligned sequence features. The spatially aligned sequence features and protein evolution-specific scoring matrix are input into the cross-attention unit in the multimodal local spatial perception module for information fusion to obtain fused protein dynamic features.

[0068] Among them, the big oracle model can be the ESM2 language model, which serves as a sequence feature extraction and tokenization model. The enhanced protein residue sequence features (i.e., protein sequence features) extracted by the ESM2 language model contain useful information such as amino acid types, functional characteristics, and evolutionary information.

[0069] To enhance the model's ability to model the protein chain structure and its rationality, this invention proposes a multimodal local spatial awareness Transformer, which aims to integrate the dynamic structural information of proteins, the sequence feature information extracted by the large language model, the protein position-specific scoring information, and the protein static two-dimensional contact map representing spatial features.

[0070] To further enhance the ability of protein conformation prediction models to perceive the basic biological structure, sequence characteristics, and temporal evolution of proteins, this application proposes a multimodal local spatial perception module with a Transformer structure. The processing of the multimodal local spatial perception module is as follows:

[0071] (1) Fusion of protein sequence features and protein spatial contact map

[0072] First, the contact distances of protein residues are calculated. These contact distances are then fused with the protein spatial contact map and converted into a Gaussian probability matrix. Among them, the Gaussian probability matrix The spatial probability relationship between the i-th residue and the j-th residue in the protein backbone in the i-th row and j-th column. The calculation method is as follows:

[0073]

[0074] Among them, map_contact i,j This represents the spatial contact relationship between the i-th and j-th residues in the protein spatial contact diagram. A value of 1 indicates that the two corresponding amino acid residues make contact at a reasonably set threshold, while a value of 0 indicates that there is almost no contact. i,j Represents the Euclidean distance between the i-th residue and the j-th residue (based on center C). α (Atomic computation); σ represents variance, which is an adjustable hyperparameter.

[0075] Then, the attention matrix attn is obtained by performing self-attention calculation based on protein sequence features. matrix Set a hyperparameter λ that can be automatically learned and adjusted for attn. matrix and After weighted calculation and softmax optimization, the final protein sequence spatial weighting matrix is ​​obtained, where the i-th residue in the protein sequence spatial weighting matrix is ​​influenced by the combined coupling effect of the j-th residue in the space_matrix. i,j Calculation method:

[0076]

[0077] in, This represents the importance of residues i and j in the attention matrix obtained by self-attention calculation of protein sequence features; E represents the number of residues in the current protein; fused represents attn. matrix and The fused weighted matrix obtained by weighted calculation i,jThis indicates the importance of the i-th residue and the j-th residue in the mixed weighting matrix.

[0078] After multiplying the protein sequence features with the final spatial weighted matrix space_matrix, self-attention calculation is performed through the self-attention unit of the multimodal local spatial awareness module to obtain the spatially aligned sequence features Align_Seq_embedding.

[0079] (2) The spatially aligned sequence features Align_Seq_embedding are fused with the protein evolution specificity scoring matrix PSSM:

[0080] This is implemented using cross-attention:

[0081] Q seq =Align_Seq_embeddingW Q

[0082] K pssm =PSSMW K

[0083] V pssm =PSSMW V

[0084]

[0085] Where PSSM represents the protein evolution specificity scoring matrix; W Q W K W V These represent the projection matrices of Query, Key, and Value during the attention calculation process for spatially aligned sequence features. Query, Key, and Value are the spatially aligned sequence features, PSSM, and PSSM, respectively; Q seq K pssm V pssm These represent the results of projection of Query, Key, and Value, respectively; the superscript T indicates transpose; CrossAttn res The fused protein dynamic features not only include features extracted from the large language model and the protein spatial contact map, but also fuse and enhance the position-specific information in the protein position-specific scoring matrix; d k This indicates the number of dimensions of the features obtained after the Value is projected, which facilitates normalization calculations.

[0086] In one embodiment, such as Figure 2As shown, the noise prediction module includes: a data information fusion layer, a first persistent spatiotemporal attention layer, a second persistent spatiotemporal attention layer, a third persistent spatiotemporal attention layer, and a fourth persistent spatiotemporal attention layer.

[0087] In one embodiment, the fused protein dynamic features, L-frame random noise, and K-frame initial protein conformation are input into a noise prediction module. Guided by the fused protein dynamic features and K-frame initial protein conformation, the noise prediction module obtains noise prediction results, including:

[0088] The L-frame random noise and K-frame initial protein conformation input data are fused together to obtain fused protein conformation information. The fused protein dynamic features and the fused protein conformation information are then processed in a first persistent spatiotemporal attention layer to obtain first protein dynamic spatiotemporal features. The first protein dynamic spatiotemporal features and the fused protein dynamic features are then processed in a second persistent spatiotemporal attention layer to obtain second protein dynamic spatiotemporal features. The second protein dynamic spatiotemporal features and the fused protein dynamic features are then processed in a third persistent spatiotemporal attention layer to obtain third protein dynamic spatiotemporal features. Finally, the third protein dynamic spatiotemporal features and the fused protein dynamic features are processed in a fourth persistent spatiotemporal attention layer to obtain the noise prediction result.

[0089] The fused protein dynamic features are obtained by the multimodal local spatial perception module of the protein conformation prediction model based on the first conditional guidance information. After the multimodal local spatial perception module obtains the fused protein dynamic features by inputting the first conditional guidance information for the first time, it can directly use the fused protein dynamic features obtained for the first time to input the noise prediction module in all subsequent predictions without further feature fusion and enhancement of the first conditional guidance information.

[0090] In the noise prediction module, when performing self-attention computation on the spatial structure, the fused protein dynamic features are used interchangeably and ultimately fused with the protein dynamic spatiotemporal features. The fusion method can be briefly described as follows: the fused protein dynamic features and the fused protein conformation information are dimensionally linked, and dimensionality reduction is performed based on a fully connected linear layer, ultimately obtaining protein dynamic spatiotemporal features that incorporate multiple biophysical structural constraints. This iterative information fusion and attention computation on the spatial structure is called persistent spatially perceptive attention.

[0091] The self-attention on the time scale (i.e., temporal self-attention) uses the standard multi-head causal self-attention of the Transformer.

[0092] The module consists of a persistent spatiotemporal self-attention layer comprised of spatial cross-attention, temporal self-attention, spatial self-attention, and a feedforward neural network. The noise prediction module contains four persistent temporal self-attention layers.

[0093] In one embodiment, the first persistent spatiotemporal attention layer, the second persistent spatiotemporal attention layer, the third persistent spatiotemporal attention layer, and the fourth persistent spatiotemporal attention layer are persistent spatiotemporal attention networks with the same structure. The persistent spatiotemporal attention network includes spatial cross attention, temporal self-attention, spatial self-attention, and a feedforward neural network connected in sequence.

[0094] Taking the processing of the first persistent spatiotemporal attention layer as an example, the fused protein dynamic features are persistently injected into the spatial cross-attention of the first persistent spatiotemporal attention layer. After spatial cross-attention calculation is performed on the fused protein conformation information input to the spatial cross-attention of the first persistent spatiotemporal attention layer, the output features are temporally embedded and then input into the temporal self-attention of the first persistent spatiotemporal attention layer for temporal self-attention calculation. After spatial self-attention calculation is performed, the features are input into the feedforward neural network of the first persistent spatiotemporal attention layer for processing to obtain the first protein dynamic spatiotemporal features.

[0095] It should be understood that the processing of the second, third, and fourth persistent spatiotemporal attention layers is the same as that of the first persistent spatiotemporal attention layer. The difference is that the input of the second, third, and fourth persistent spatiotemporal attention layers is the fused protein dynamic features of the previous layer and the continuously injected data.

[0096] In one embodiment, the protein conformation prediction model is trained as follows:

[0097] Molecular dynamics simulation trajectory data of different proteins with 77 or fewer residues within 10 ns under physiological conditions and molecular dynamics simulation trajectory data of different proteins with 77 to 200 residues were obtained and preprocessed to obtain multiple sample datasets, each of which includes L frames of molecular dynamics simulation trajectory data. The molecular dynamics simulation trajectory data of each sample dataset from frame K+1 to frame L were masked to obtain the masked sample dataset, which includes the masked sample data corresponding to each sample dataset. The sample dataset, the masked sample dataset, and the first condition guidance information were input into the protein conformation prediction model to be trained for training to obtain the trained protein conformation prediction model. L frames of Gaussian noise were randomly sampled, and the L frames of Gaussian noise, the first condition guidance information, and the third condition guidance information were input into the trained protein conformation prediction model to predict conformational change trajectories to obtain the protein conformational change trajectory prediction results. Based on the protein conformational change trajectory prediction results, the network parameters of the trained protein conformation prediction model were fine-tuned using a reinforcement learning fine-tuning algorithm optimized by protein conformational dynamics strategy to obtain the trained protein conformation prediction model.

[0098] The masked sample data is obtained by masking the molecular dynamics simulation trajectory data of frames K+1 to L in each sample data using a mask.

[0099] The third conditional guidance information can be either masked sample data or other new K-frame protein initial conformations.

[0100] Among these methods, molecular dynamics (MD) simulation trajectory data of proteins are collected and organized as the basis for the learning of the protein conformation prediction model to be trained.

[0101] In one example, a molecular dynamics simulator based on a quantized resonance model was used to simulate the molecular dynamics trajectory data of proteins with 77 or fewer residues under physiological conditions within a 10 ns range, using femtosecond simulation technology (10e-15 s). For molecular dynamics trajectory data of large and diverse proteins with more than 77 residues up to 200, the publicly available dataset DynamicPDB was selected and used. Since the simulation times for these proteins in the DynamicPDB dataset may be long, subsequent normalization processing is required. The static PDB structure files for all proteins were obtained from the PDB database.

[0102] The preprocessing of protein molecular dynamics simulation trajectory data includes coordinate transformation, which converts the protein's traditional Cartesian coordinates to internal coordinates, i.e., from the main chain... The protein is represented by a 7-tuple consisting of seven dihedral angles (torsion angles) along with the side chains (χ1, χ2, χ3, χ4). This representation more fundamentally describes the flexibility of the protein and reduces redundancy in rotation and translation. Each torsion angle is further encoded as sin and cos values ​​to handle the periodicity of the angle.

[0103] In one example, preprocessing includes data cleaning and preliminary processing, as well as reading and preprocessing the molecular dynamics simulation trajectory data of the protein to obtain data for each sample.

[0104] Data organization and preliminary processing include:

[0105] (1) Organize and process all protein molecular dynamics simulation trajectory data according to the standard format, and select and retain the molecular dynamics simulation trajectory data of proteins in .dcd format. For the molecular dynamics simulation trajectory data of proteins obtained from DynamicPDB, the molecular dynamics simulation trajectory data of proteins are segmented according to the time sequence, and the simulation duration of each segment is kept at 10ns. The data is renamed according to the protein code and sequence number to facilitate random selection in the subsequent training phase. (2) For the molecular dynamics simulation trajectory data obtained for each protein, process it into .npy files using the numpy library of Python for easy saving. (3) Package and build the dataset.

[0106] The reading and preprocessing of protein molecular dynamics simulation trajectory data includes:

[0107] (1) Read the molecular dynamics simulation trajectory data of the protein and set the reading interval (the default is 1, which means 1 femtosecond) to complete the equal-interval sampling process of the molecular dynamics simulation trajectory data of the protein and obtain L frames of molecular dynamics simulation trajectory data of the protein.

[0108] (2) Read the molecular dynamics simulation trajectory data of the sampled protein as vector data (using the MdTraj library), filter out H atoms and retain key heavy atoms, and perform atomic grouping and normalization according to residues, that is, the shape is transformed from the original (l,n_atoms,3) to (l,n_residues,14,3), where l represents the time length of the currently read molecular dynamics data, and the frame interval depends on the time interval set in the sampling process; n_atoms is the number of heavy atoms (non-hydrogen atoms) in the protein's composition structure; n_residues represents the number of residues in the protein; the dimension of 3 represents the Cartesian coordinates (x,y,z) of the atom in three-dimensional Euclidean space; the dimension of 14 represents the minimum dimension required to accommodate all the key heavy atom coordinates of any one of the 20 standard amino acids without loss. In this way, all 20 residues can be uniformly represented as a 14×3 coordinate matrix.

[0109] Each protein conformation is converted from a standard Cartesian coordinate representation to an Atom14 standard Cartesian coordinate representation (a coordinate representation method used in most studies that use deep learning to represent protein structures), such as... Figure 3 The diagram shows a standard representation of Atom14 atoms in a protein, where key atoms are grouped and filled according to residues.

[0110] Atom14 indicates that each residue group will contain:

[0111] Main chain heavy atoms (4): N, C α Figure 4 shows a schematic diagram of the protein backbone twist angle, C, O; where N (Amide Nitrogen): this atom is the nitrogen atom in the amino (-NH2) group, covalently linked to the carbonyl carbon (C) of the preceding amino acid residue via a peptide bond; C α (α-Carbon): The central chiral carbon atom of each amino acid residue. C (Carbonyl Carbon): This atom is the carbon atom in the carboxyl group (-COOH), which is connected to the carbonyl oxygen (O) via a double bond, forming a peptide bond that links to the next amino acid residue. O (Carbonyl Oxygen): This atom is connected to the carbonyl carbon (C) via a double bond, forming the carbonyl functional group (C=O).

[0112] Key side-chain heavy atoms (10): CB, CG, CD, CE, NZ, CG2, CD2, CE2, NE2, OXT; First, the naming rules for these 10 key side-chain heavy atom sites are as follows: Greek letter sequence: Starting with the first side-chain atom immediately adjacent to the main chain α-carbon (Cα), the Greek letters Beta (β), Gamma (γ), Delta (δ), Epsilon (ε), and Zeta (ζ) are used to mark its sequential position on the side chain. The letter before the atom symbol represents its position; for example, CB represents the carbon atom at the Beta position. Branch number: When the side chain branches, atoms at the same Greek letter position but on different branches will be distinguished by a numerical suffix to ensure the uniqueness of the naming.

[0113] Therefore, these ten key side-chain heavy atom sites are represented as follows:

[0114] CB (Cβ, Beta-Carbon): β-carbon. The first carbon atom on the side chain of all amino acids except glycine, which is directly covalently linked to the Cα carbon of the main chain.

[0115] CG (Cγ), CD (Cδ), CE (Cε): These represent the γ-carbon (Gamma-Carbon), δ-carbon (Delta-Carbon), and ε-carbon (Epsilon-Carbon), respectively. They indicate the carbon atoms at the second, third, and fourth positions from the α-carbon in the unbranched (or main branched) chain. For example, in the side chain of methionine, the carbon atom sequence is CB-CG-CD-CE.

[0116] NZ (Nζ, Zeta-Nitrogen): ζ-nitrogen. This specifically refers to the nitrogen atom at the sixth (Zeta) position, the terminal nitrogen atom in the lysine side chain. It is the functional amino group (-NH3) in the lysine side chain. + The core of ).

[0117] CG2, CD2, CE2, NE2: These atoms with numerical suffixes indicate that they are located on branches of the side chain: CG2: Usually refers to the second carbon atom in the Gamma position of amino acids that branch from the β-carbon (such as valine and isoleucine). CD2: Refers to the second carbon atom in the Delta position of amino acids that branch from the γ-carbon (such as leucine). CE2 refers to the second carbon atom in the Epsilon position on the benzene ring (if present) of some amino acids (such as phenylalanine).

[0118] NE2: Specifically refers to the second nitrogen atom in the Epsilon position on the imidazole ring of histidine.

[0119] OXT (Terminal Oxygen): C-terminal oxygen atom. This is a very special atom that exists only on the last amino acid residue of the entire protein chain.

[0120] For each amino acid residue, the missing atom is represented using a mask_origin. A value of False at each position in mask_origin indicates that the heavy atom is missing from the composition of the current residue and will be ignored in subsequent neural network feature extraction, mask co-embedding, and guided diffusion calculations. A value of True at each position indicates that the heavy atom is present in the composition of the current residue.

[0121] (3) However, for the input of subsequent protein conformation prediction models, protein conformations represented by Cartesian coordinates cannot be directly input. Instead, the protein conformation should be converted to a torsion angle representation, such as... Figure 5 The seven-element torsion angle of the protein residues is shown.

[0122] The specific conversion process is as follows:

[0123]

[0124] Where, pre-ω, ψ, χ1, χ2, χ3, χ4 represent seven types of torsion angles. The first three are the torsion angles formed by the Cα of a residue (not at the beginning or end) on the main chain with the N atom of the preceding peptide bond, the C atom of the subsequent peptide bond, and the Cα of the subsequent residue. The remaining four are the torsion angles that the side chain may form (the side chain state of each amino acid residue may be different). Let any twist angle be represented by a tuple after the transformation; It represents the set of real numbers.

[0125] Finally, for each torsion angle, it can be represented based on sine and cosine encoding:

[0126]

[0127] Here, θ represents the set of twist angles for the entire protein, which is represented by tuples of twist angles for the corresponding number of residues. The set is mainly represented in matrix form; This represents any torsion angle after sine and cosine encoding. That is, sine and cosine encoding is performed on any torsion angle tuple representation in θ to obtain the corresponding encoded torsion angle representation.

[0128] Note: For this seven-element torsion angle representation, a mask 2 also needs to be constructed (distinguished from the mask 1 in the Atom14 representation). Each bit value is True (existence), indicating that this type of torsion angle exists within the current residue range. Setting each bit value to False (absence) means that this type of torsion angle does not exist within the current residue range and should be ignored in subsequent calculations and format conversions.

[0129] To ensure the efficiency of the protein dynamic conformation set generation process, the system adopts a non-autoregressive inference framework. To ensure that the protein conformation prediction model can accurately predict and generate temporally coherent protein dynamic conformations (i.e., white matter conformational change trajectories) under this framework based on given initial conditions, this application employs a conditional denoising process training mode for protein conformation dynamics inference.

[0130] Among them, such as Figure 6 As shown, the training phase of the protein conformation prediction model consists of forward noise addition and backward noise reduction modes. The masked sample data and the first conditional guidance information are used as guiding conditions. During the forward noise addition process, based on the sample data... By performing a forward noise addition process at each time step, until noise is added to the T-th time step, the noisy sample data is obtained. The noisy sample data This can be viewed as Gaussian noise. In the reverse denoising process, based on the noisy sample data... By denoising at each time step, the final sample data can be obtained. Therefore, through training, protein conformation prediction models can learn to denoise input random noise based on conditional guidance, thereby predicting the trajectory of protein conformational changes.

[0131] The denoising process at each diffusion time step in the prediction phase, based on the multimodal local spatial perception module, noise prediction module, and denoising module, is as follows: Figure 7 As shown:

[0132] The protein residue sequence is input into the large oracle model of the multimodal local spatial perception module for feature extraction to obtain protein sequence features. The protein sequence features and the protein spatial contact map are then fused by the feature fusion unit of the multimodal local spatial perception module, and self-attention calculation is performed by the self-attention unit of the multimodal local spatial perception module to obtain spatially aligned sequence features. The spatially aligned sequence features and the protein evolution-specific scoring matrix are input into the cross-attention unit of the multimodal local spatial perception module for information fusion to obtain fused protein dynamic features.

[0133] Taking the denoising process at time step t as an example: the fused protein dynamic features, the masked sample data in the masked sample dataset, and the L-frame protein conformation x to be denoised at time step t. t The input noise prediction module predicts noise, which is then output as the noise prediction output. Based on the predicted noise, the L-frame protein conformation x to be denoised at time step t is determined. t Denoising is performed to obtain the L-frame protein conformation x at time step t after denoising. t-1 .

[0134] The noise prediction module is treated as a conditional denoising autoencoder. This conditional denoising autoencoder is trained using the following two training modes:

[0135] (1) Single-frame guided multi-frame trajectory generation mode

[0136] Objective: This mode aims to train a protein conformation prediction model to learn a mapping ability, namely, to receive a single initial conformation frame of a protein as input and, based on this, generate a complete protein conformational change trajectory containing L frames. Training Implementation Method:

[0137] During the training phase, a complete protein conformational change trajectory T = {F1, F2, ..., FL} containing L frames is taken from the sample dataset. The first frame protein conformation F1 (i.e., K = 1) is used as one type of conditional information, and a mask vector M that exposes only the first frame is constructed. In each step of the denoising training, the protein conformation prediction model adds noise to the input sample data (i.e., the protein conformational change trajectory of L frames). The noisy sample data Tt, the current time step t, the first conditional guidance information, the masked sample data exposing the first frame protein conformation F1, and the mask vector M are provided to the noise prediction module of the protein conformation prediction model. The architecture of this noise prediction module enables it to simultaneously process the complete noisy sample data, the first conditional guidance information, the masked sample data exposing the first frame protein conformation F1, and the mask vector M. This process is guided by the existence of the first frame protein conformation F1, enabling the noise prediction module to learn to predict the noise distribution in subsequent denoising processes and allowing the protein conformation prediction model to generate subsequent trajectories based on the first frame protein conformation F1.

[0138] (2) Multi-frame guided trajectory continuation prediction mode

[0139] Objective: This mode aims to train a protein conformation prediction model to grasp more complex dynamic laws. Specifically, it takes an initial protein conformation fragment containing K frames (i.e., L>K>1) as input and, based on the spatiotemporal dynamic information contained in this fragment, guides and predicts the protein conformational change trajectory that generates L frames. Training Implementation Method:

[0140] The training process for this mode is similar to that of Mode 1, except that one type of conditional information is the protein conformations of the first K frames {F1, F2, ..., FK}, and the mask vector M exposes these first K frame protein conformations. The noise prediction module receives a combination of inputs including the noisy sample data Tt, the time step t, the first conditional guidance information, and the masked sample data exposing the first K frame protein conformations and their corresponding masks. In this mode, the persistent spatiotemporal attention mechanism within the noise prediction module is specifically designed to effectively process the unmasked first K frame protein conformations in its encoding path, thereby analyzing and extracting spatiotemporal features containing dynamic trends. These spatiotemporal features then work together with the original protein spatial structure information to more strongly guide the denoising process, ensuring that the predicted subsequent trajectories conform to the dynamic laws exhibited by the initial segment.

[0141] It should be understood that by designing the noise prediction module as a conditional denoising autoencoder and training it using a mask-guided strategy, the core objective of efficiently and accurately predicting and generating long-term protein conformational change processes from a given initial state is achieved.

[0142] The protein conformation prediction model is based on a conditional denoising diffusion framework and employs a deterministic sampling strategy for efficient trajectory sampling. The training and sampling process of the protein conformation prediction model will be described in detail below.

[0143] (1) Training of the protein conformation prediction model, the training process is as follows:

[0144] ① Add noise

[0145] (Continuous time sampling)

[0146] (Standard Gaussian noise)

[0147]

[0148] in, Time step parameters, Represents a continuous uniform distribution from 0 to 1, and t represents the time step parameter obtained by random sampling from this distribution; Note: Discrete time step sampling is not used in actual implementation here, because continuous time sampling is more convenient in subsequent related calculations; Noise parameters, Represents standard Gaussian noise (mean 0, covariance matrix is ​​identity matrix I), ∈ α This represents the true noise distribution obtained by randomly sampling from the standard Gaussian noise distribution; This represents the molecular dynamics simulation trajectory data of the original, noise-free protein, where B represents the batch size (the number of samples in a single training iteration), L represents the time dimension (L frames of sample data), and NUM... RES d represents the number of residues in the corresponding protein, and d represents the number of feature dimensions. Represents the cumulative noise dispatch coefficient; The set of multi-frame protein dynamic conformations after noise addition at time step t during the noise addition process.

[0149] ② Prediction noise. Therefore, the prediction noise of the protein conformation prediction model at time step t during the training step can be described as follows:

[0150]

[0151] Where 'c' represents the masked sample data; 'LLM' represents the large language model; 'Seq' represents the protein residue sequence; 'MLAST' represents the multimodal spatial sensing module; and 'Model' represents the model. θ θ represents the protein conformation prediction model, where θ is the network parameter of the protein conformation prediction model; Contact represents the protein spatial contact map; MLAST(PPSM,Contact,LLM(Seq)) represents the dynamic features of the fused protein; c_guide represents the input guidance information (MLAST(PPSM,Contact,LLM(Seq)),c), which is represented in tuple form. The predicted noise distribution (i.e., the noise prediction result) is because, in the actual model architecture implementation, the Model... θ The parameters predicted are the noise distribution parameters, namely the mean and variance.

[0152] It should be understood that during the training phase, the input guidance information includes the first guidance information and the masked sample data (i.e., the masked sample data is used as the initial conformation of the K-frame protein during training, serving as a guide); during the prediction fine-tuning process based on the trained protein conformation prediction model, the input guidance information includes the first guidance information and the third guidance information, where the third guidance information can be the masked sample data or other initial conformations of the K-frame protein; in the practical application phase of the trained protein conformation prediction model, the input guidance information includes the first guidance information and the second guidance information, where the second guidance information specifies the initial conformation of the K-frame protein of interest for researchers or users, thereby predicting the protein conformation change trajectory corresponding to the initial conformation of the K-frame protein of interest.

[0153] ③ Training objectives of the protein conformation prediction model: The training objectives of the protein conformation prediction model during the training phase are as follows:

[0154]

[0155] in, The loss function during the training phase. For expectation, ∈ α This represents the actual noise distribution during the noise addition process. For the sample dataset, x t Let L be the protein conformation of the frame to be denoised at time step t, where t represents a continuous uniform distribution from 0 to 1. The time step for uniform sampling is defined by c_guide, which represents the input guiding condition information. θ represents the predicted noise distribution, and θ represents the network parameters of the protein conformation prediction model.

[0156] (2) Sampling process

[0157] 1. Denoising sampling allows the denoising diffusion process, which occurs when predicting protein conformational dynamics, to proceed across time steps:

[0158] Original Trajectory Estimation: First, the noise distribution predicted at any time step t is used. Molecular dynamics simulation trajectory data of the original noiseless protein were directly estimated using algebraic transformations. The optimal prediction is denoted as Furthermore, for ease of subsequent explanation, use A noise vector representing the current state.

[0159]

[0160] The main purpose of the original trajectory estimation is to make the sampling process more robust and allow for large step jumps.

[0161] 2. Deterministic Reverse State Transition Denoising:

[0162] In obtaining the estimate of the original trajectory Then, the state x at the previous moment is calculated using a deterministic formula. t-1 This formula linearly combines the protein conformation pointing to the final predicted target with the noise vector pointing to the current state. Therefore, from x... t To x t-1 The noise reduction process is as follows:

[0163]

[0164] in:

[0165] t = [T, T-Δt, T-2Δt, ...] (Time steps can be sampled in multiple steps)

[0166] Where Δt represents the time step spanned; t is the time step that can be sampled across (here it is still a continuous time step); and T is the maximum time step.

[0167] 3. Iterative denoising:

[0168] The deterministic reverse state transition denoising method performs iterative denoising across time steps and finally obtains a reasonable protein conformational change process through condition-guided sampling (visualizing and analyzing the generated set of dynamic protein conformations). This sampling method has advantages such as determinism, reversibility and faster sampling speed, and is particularly suitable for protein conformational dynamics prediction tasks that require precise control of the generation process.

[0169] The trained protein conformation prediction model already possesses the ability to generate temporally consistent trajectories. However, to ensure it more strictly adheres to physical laws, a reinforcement learning fine-tuning algorithm for protein conformation dynamics policy optimization is designed to adjust the network parameters of the trained model (i.e., the fine-tuning phase). This reinforcement learning fine-tuning algorithm for protein conformation dynamics policy optimization includes the construction of a comprehensive reward function and the Protein Conformational Dynamics Policy Gradient Optimization (PCDPO) algorithm.

[0170] The comprehensive reward function is a multi-faceted reward function incorporating multiple physical, geometric, and biological constraints. This function evaluates the rationality of the generated conformations in real time, including: an energy reward (Renergy): the total potential energy calculated using tools such as OpenMM, encouraging low-energy conformations; a collision detection reward (Rclash): penalizing spatial collisions between atoms; and a Ramachandra reward: ensuring dihedral angles on the main chain. Within a reasonable range; Secondary structure preservation reward (Rsecondary): encourages maintaining reasonable α-helical and β-fold structures; Conformation diversity reward (Rdiversity): encourages the generation of diverse conformations and avoids mode collapse.

[0171] Among them, the protein conformation dynamics policy gradient optimization algorithm directly uses the above-mentioned comprehensive reward function for reinforcement learning and policy gradient updates. In the high-dimensional and complex optimization scenario of protein conformation, this is prone to training instability or "catastrophic forgetting" (i.e., the protein conformation prediction model easily forgets the basic dynamic laws learned from molecular dynamics simulation trajectory data). To address this, a novel protein conformation dynamics policy optimization algorithm is designed for the analysis and prediction model of protein dynamic conformation changes.

[0172] In the fine-tuning phase, the trained protein conformation prediction model is treated as a "policy network," where the "action" is the selection of an evolutionary direction at each denoising step within the trained model. An independent neural network model, the value network, is constructed. This value network evaluates the long-term expected cumulative reward (i.e., "state value") obtained by following the current policy in the current protein conformation state. To ensure proper backpropagation during backpropagation, the value network is a branch network extended from the pruned protein conformation prediction model. The conformational advantage function A(s,a) is calculated by comparing the "real" single-step reward with the value network's expectation. This function characterizes how much better a specific "action" a (denoising direction) is than the average value in the current protein dynamic conformation trajectory state s, indicating a more effective direction for policy updates. To achieve stable updates, a pruning target loss function with a probability ratio pruning mechanism is employed. The core idea behind this pruning target loss function is to limit the deviation between the old and new policies within a single update step. This effectively prevents drastic strategy changes caused by an excessively strong reward signal, ensuring that the protein conformation prediction model, while learning to follow physical laws, does not destroy the basic protein dynamics knowledge learned during the training phase. Simultaneously, it achieves stable fine-tuning by embedding static protein structure and other biochemical rationality assessment indicators into dynamic process analysis and prediction without introducing additional molecular dynamics simulation trajectory data, fundamentally resolving the core contradiction between "temporal continuity reasoning of conformational dynamics" and "physical geometric rationality."

[0173] In summary, the protein conformation dynamics policy gradient optimization algorithm proposed in this application evaluates the long-term value of the current state and calculates the conformation advantage function by using an independent conformation value network capable of gradient calculation and propagation. Through a clipping objective loss function with a clipping mechanism, the network parameters of the policy network (protein conformation prediction model) are stably updated, optimizing it towards directions that yield higher overall rewards, without forgetting the fundamental principles learned from MD data (i.e., molecular dynamics simulation trajectory data).

[0174] The optimized torsion trajectory generated by the protein conformation prediction model is inversely converted back to the Atom14 standard Cartesian coordinates of each residue using the Natural Extension Frame of Reference (NeRF) algorithm, and can be further refined into an all-atom structure using tools such as Pyrosetta.

[0175] In one embodiment, the training objective of the protein conformation prediction model to be trained is:

[0176]

[0177] in, The loss function during the training phase. For expectation, ∈ α This represents the actual noise distribution during the noise addition process. For the sample dataset, x t Let L be the protein conformation of the frame to be denoised at time step t, where t represents the time step from the time step t. The time step for uniform sampling is defined by c_guide, which represents the input guiding condition information. θ represents the predicted noise distribution, and θ represents the network parameters of the protein conformation prediction model.

[0178] In one embodiment, the reinforcement learning fine-tuning algorithm for optimizing protein conformational dynamics strategies includes: constructing a comprehensive reward function that incorporates physical, geometric, and biological constraints based on energy rewards, conflict detection rewards, Ramachandran rewards, secondary structure preservation rewards, and conformational diversity rewards;

[0179] Based on the probability ratio of single-step conformational transitions in the trained protein conformation prediction model, a pruning objective loss function is constructed. Based on the comprehensive reward function, an auxiliary objective loss function is constructed using the conformational value network. Based on the pruning objective loss function and the auxiliary objective loss function, an overall optimization objective function is constructed. Based on the overall optimization objective function, the network parameters of the trained protein conformation prediction model and the network parameters of the conformational value network are updated collaboratively until the performance of the trained protein conformation prediction model converges, thus obtaining a well-trained protein conformation prediction model.

[0180] In one embodiment, the overall optimization objective function for:

[0181]

[0182] in, To prune the target loss function, λ V The weighting coefficients of the loss term in the auxiliary objective loss function, To assist the target loss function, λ KL Let be the weight coefficients of the KL divergence regularization term. Here, φ is a KL divergence regularization term, φ is the network parameter of the conformational value network, and θ is the network parameter of the protein conformation prediction model.

[0183] In one embodiment, the auxiliary target loss function is:

[0184]

[0185] Among them, V φ (x T (c_guide) represents the value reward score for the conformational value network assessment, x T L frames of Gaussian noise, c_guide is the input guidance condition information, R(x0) is the comprehensive reward function, and proein~S is the protein sequence obtained by randomly sampling from the protein sequence library S with different residue sequences.

[0186] In one embodiment, the pruning target loss function is:

[0187]

[0188] Where T is the maximum time step, r t (θ) represents the probability ratio of single-step protein conformational transfer, A t Let clip_b be the conformational dominance function at time step t, and clip_b be a hyperparameter defining the clipping boundary; clip(·) represents clipping.

[0189] Another core contribution of this application is the design of a novel reinforcement learning fine-tuning algorithm for optimizing protein conformation dynamics strategies for the proposed protein conformation prediction model. This reinforcement learning fine-tuning algorithm for optimizing protein conformation dynamics strategies can achieve better weights (i.e., network parameters) based on the network parameters obtained during the conditionally guided training phase (mainly the multimodal local spatial awareness module and the noise prediction module) and on physical, geometric, and biological constraints without requiring additional data.

[0190] (1) Design a comprehensive reward function that incorporates physical, geometric, and biological constraints:

[0191] R(x0) = w1·R energy (x0)+w2·R clash (x0)+w3·R rama (x0)+w4·R secondary (x0)+w5·R diversity (x0)

[0192] Where R(x0) is the comprehensive reward function, x0 represents the multi-frame protein conformation change trajectory generated by the protein conformation prediction model, and w1, w2, w3, w4, and w5 represent the weights corresponding to each reward component.

[0193] The specific calculation methods for each reward component are as follows:

[0194] I. Energy Reward R energy :

[0195] Real-time calculation of total potential energy based on the Python interface of OpenMM:

[0196]

[0197] Among them, E total This includes bond length, bond angle, dihedral angle, electrostatics, and van der Waals interaction energy; k B T: Energy normalization parameter, i.e., Boltzmann constant k B Multiply by temperature T 温 (Under physiological conditions); E ref : The energy required for this protein to maintain a stable conformation under physiological conditions (obtained directly via OpenMM).

[0198] II. Conflict Detection Reward R clash :

[0199]

[0200] Where, r ij Let r be the distance between the i-th and j-th atoms of the protein. vdw Let λ be the sum of the van der Waals radii of all atoms within the protein. clash To extract the hyperparameters of the collision detection weights, r thresh The set interatomic distance collision detection threshold (unit: r) ij Consistent), σ clash For variance; This means that calculations are only performed when a conflict is detected.

[0201] III. Ramachandran Reward R rama :

[0202] Based on amino acid-specific dihedral distribution:

[0203]

[0204] Among them, P rama The Ramachandran distribution obtained from PDB statistics can be directly analyzed using the MdTraj library. φ i ,ψ i There are two types of dihedral angles. aa i Let represent the i-th residue. E represents the number of residues in the current protein; the reason it is 2 to L-1 in the formula is because the first and last residues do not have a complete φ. i ,ψ i horn.

[0205] IV. Secondary Structure Retention Bonus R secondary :

[0206]

[0207] Among them, s i Secondary structure types (α-helix, β-fold, random coil); For the predicted results, For reference only; Indicates the number of secondary structure types; w ss This represents the proportion of the i-th secondary structure in the overall secondary structure reward calculation; the secondary structure retention reward can be directly obtained through the Pyrosetta interface.

[0208] V. Construct Diversity Reward R diversity :

[0209]

[0210] Where j represents the j-th generation result selected from the historical generation record for diversity reward calculation, and RMSD(·,·) represents the root mean square deviation calculation. It represents the trajectory of conformational changes of representative proteins throughout history.

[0211] For proteins with the same residue sequence, the protein conformation prediction model will be fine-tuned through multiple rounds of reinforcement learning in order to promote the generation of diverse conformations and avoid pattern collapse.

[0212] Therefore, the protein conformation prediction model sampling is achieved by guiding the sampling to generate a set of protein conformation change trajectories x0, and then comparing these trajectories with historically representative protein conformation change trajectories. Sampling is performed on a time scale, where j represents the j-th generation result selected from the historical generation record for diversity reward calculation.

[0213] The above rewards can be calculated in real time based on protein software packages such as OpenMM, Pyrosetta, and MdTraj, hidden Markov models, PCA analysis, or physical formulas.

[0214] (2) Reinforcement learning fine-tuning algorithm for protein conformation prediction model

[0215] I. The complete diffusion denoising trajectory τ can be redefined:

[0216] τ=(x t ,a T-1 ,x T-1 ,...,a0,x0)

[0217] Among them, a t For time step t, denoise the action, t∈0,1,2…T; x0 represents the predicted protein conformational change trajectory; x T-1 x represents the Gaussian noise of frame L.T The result after denoising at time step T.

[0218] II. Pruning objective loss function used in reinforcement learning fine-tuning algorithm for protein conformation dynamics strategy optimization:

[0219] This application designs a policy gradient estimation algorithm suitable for protein conformation dynamics prediction tasks. In order to solve the training instability in the high-dimensional continuous space of protein conformation dynamics analysis and prediction, the reinforcement learning fine-tuning algorithm has designed a clipping mechanism to ensure the robustness of policy updates.

[0220] First, define the probability ratio r of a single-step conformational transition. t (θ), which measures the protein conformation change prediction agent model with new network parameters. θ Compared to a model that uses sampled data and preserves previous parameters, multi-frame conformation denoising (from x) T To x T-1 Comparison of evolutionary directions / strategies in the process:

[0221] In order to convert the Model during the training phase θ To distinguish this, π will be used from this point onward. θ Refers to the protein conformation prediction model. θ (Also known as the policy network), and its network parameters θ (network parameters include weights and parameters) can determine the corresponding policy and decision mode at the moment (i.e., determine which direction the configuration will change next).

[0222]

[0223] Among them, logr t (θ): In the derivation of the formula, for ease of representation and derivation, the probability r of the denoising strategy is calculated. t (θ) represents the likelihood; σ is the fixed variance, calculated based on the relevant parameters of the forward noise addition process; ∈ represents the noise distribution predicted by the protein conformation prediction model obtained during the training phase at the current time step; This represents the noise distribution predicted at time step t by the strategy of the protein conformation prediction model before network parameter updates; ∈ θ This represents the noise distribution predicted at time step t by the strategy of the conformational change prediction model based on network parameter updates.

[0224] To prevent the conformational evolution strategy from collapsing due to excessively large single update steps, policy updates are limited to a reliable range. The final pruning objective loss function... The definition is as follows:

[0225]

[0226] Among them, A t is the conformational advantage function at time step t, which indicates how much better a specific evolutionary action is than the average level in the current conformational state; T represents the maximum time step, which is written in discrete time step form for convenience, but continuous time step sampling is still used in actual implementation; proein~S represents the protein sequence obtained by randomly sampling from a protein sequence library S with different residue sequences; clip_b is a hyperparameter defining the clipping boundary, and clip(·) represents clipping.

[0227] This pruning objective loss function ensures that the update process of the conformational evolution strategy is smooth and stable, avoiding overly aggressive and physically incompatible explorations of energy geomorphology.

[0228] III. Construct rationality assessment network and dominance estimation

[0229] A t =|R(x) t )-R mean |

[0230] This formula intuitively represents: the denoising strategy adopted by the network parameters corresponding to the current protein conformation prediction model, and the L-frame protein conformation x to be denoised at time step t. t The reward R(x) for denoising prediction t The weighted reward calculated after storage during the previous rounds of fine-tuning training) The second normal form.

[0231] To ensure the conformational dominance function A t To participate in policy gradient calculation and optimization, the reinforcement learning fine-tuning algorithm designed in this application requires the introduction of an independent Conformational Value Network (VVN) to guide gradients during the reinforcement learning fine-tuning of the protein conformation prediction model. Here, the VVN is referred to as V. φ The network parameter of the conformational value network is φ. This conformational value network is based on the noise prediction module within the protein conformation prediction model, and is specifically used to predict the L frames of protein conformations x to be denoised at a given time step t. t Given the input guidance information c_guide, the expected value of the total future biophysical reward that can be obtained.

[0232] Therefore, the branched construct value network will be trained synchronously as a branch network by minimizing the mean squared error (MSE) during the fine-tuning training of reinforcement learning:

[0233]

[0234] Among them, V φ (X T (c_guide) represents the set of multi-frame protein conformational change trajectories obtained by the policy network in the initial time step after making decisions and denoising. This set, along with all guiding information, is then fed into the branch conformational value network for forward propagation to obtain the evaluated value reward score. Since the conformational value network shares the main network layers and network parameters with the policy network, its first parameter is also X. T , representing the trajectory state of the initial multi-frame protein conformation set to be denoised (i.e., L-frame Gaussian noise) sampled from a normal Gaussian distribution; This is the auxiliary target loss function.

[0235] IV. Learning Objectives in the Fine-tuning Phase of Reinforcement Learning

[0236] 1) The overall optimization objectives of PCDPO are as follows:

[0237] To balance policy optimization, value function fitting, and deviation from the trained protein conformation prediction model, the overall optimization objective of PCDPO is a composite loss function that includes multiple regularization terms:

[0238]

[0239] in, The overall optimization objective function; θ represents the target loss function for pruning; θ represents the network parameters of the policy network (i.e., the weights and related parameters of the protein conformation prediction model, which determine the policy optimization mode and direction of the multi-frame protein conformation set in the denoising generation process). φ represents the training objective (i.e., the auxiliary objective loss function) of the conformational value function; φ represents the weights of the conformational value network. It is a KL divergence regularization term used to penalize the new policy π. θ Compared with the trained protein conformation prediction model (or the old strategy) The significant deviation between λ and MD data ensures that the protein conformation prediction model does not forget the fundamental physical laws learned from MD data while learning new knowledge; V and λ KL These are the weighting coefficients of their respective loss terms.

[0240] 2) Algorithm flowchart:

[0241]

[0242]

[0243]

[0244]

[0245] (3) Explanation of the reinforcement learning fine-tuning algorithm process

[0246] Objective: To fine-tune the trained protein conformation prediction model to optimize the generated conformational ensemble and make it more consistent with physical, chemical and biological constraints.

[0247] I. Initialization: Policy Network: π θ Its internal network parameters θ are initialized by the network parameters of the protein conformation prediction model trained on MD data. In the description of reinforcement learning fine-tuning algorithms, it can also directly refer to its corresponding policy and decision-making mode. Conformation Value Network: V φ Its internal network parameters φ, and its network structure are mainly achieved through pruning and branch extension of the trained protein conformation prediction model network. Therefore, the value network not only shares parameters with some layers of the policy network, but also constructs an additional value score prediction branch based on the noise prediction network of the policy network (the parameters of this part of the network layer are randomly initialized). In the description of reinforcement learning fine-tuning algorithms, it can also directly refer to its corresponding policy and decision-making mode. Experience Replay Buffer: Initializes an empty buffer. Used to store the generated conformational evolution path. Guidance Condition: Provides the protein residue sequence, protein spatial contact map, protein evolution-specific scoring matrix, and K-frame initial protein conformations as the guidance condition c_guide. Hyperparameters are set: policy learning rate α_rate; value learning rate β_rate; hyperparameter clip_b defining the clipping boundary; and weight coefficients λ for the auxiliary objective loss function loss term. V The weight coefficient λ of a KL divergence regularization term. KL Total number of training iterations N iter The number of rollouts generated in each iteration cycle (N) rollout ; Epochs in each iteration cycle (D); Minibatch size (H).

[0248] This stage is the system construction process before the reinforcement learning fine-tuning algorithm is launched, aiming to lay the foundation for subsequent adaptive optimization.

[0249] Establishing the core of policy and value assessment: The reinforcement learning fine-tuning algorithm is applied to a protein conformation prediction model trained on MD data. This trained protein conformation prediction model constitutes the policy network π. θ It forms the foundation, encapsulating the basic understanding of protein conformational space. It introduces a specially designed conformational value network, V... φ Its unique function lies in proactively assessing the long-term biophysical value of any dynamic conformation throughout the entire evolutionary process. Initial guidance and execution parameters are set: a prediction process with a clearly defined physical starting point is initiated by receiving the user-provided initial protein conformation as one of the guiding conditions. Simultaneously, a reinforcement learning fine-tuning algorithm initializes a dedicated experience replay buffer. A series of operating hyperparameters are set, which together constitute a unique, complete, and controllable optimization process.

[0250] II. Reinforcement Learning Fine-tuning Algorithm Main Loop

[0251] The core of the reinforcement learning fine-tuning algorithm is a self-evolving iterative optimization loop. It drives the continuous improvement of the performance of protein conformation prediction models through the unique alternation of "exploratory sampling" and "policy value self-learning".

[0252] Step 1: Explorative Rollout Phase: This is the data generation and exploration phase. At the beginning of each iteration, the reinforcement learning fine-tuning algorithm utilizes the policy of its current version of the policy network. The model actively explores conformational space, generating a batch of novel protein conformational change trajectories that follow the laws of physical evolution. These trajectories are not repetitions of existing data, but rather the best predictions made by the protein conformation prediction model based on the current level of understanding. A comprehensive reward function R quantifies the final outcome of each protein conformational change trajectory, transforming complex physicochemical rules into a clear signal that can be optimized. All the discovered protein conformational change trajectories and their reward evaluations are stored as valuable "experience" in a buffer.

[0253] Step 2: Policy Refinement & Value Learning Phase: This is the core learning and self-evolution phase. The protein conformation prediction model will learn from the "experience" explored in the previous phase to refine its policy and improve its value judgment capabilities. Experience data sampling: from the buffer... A small batch of empirical data is extracted to drive this learning process. Core innovative computation: For each empirical trajectory, its unique analysis and computation process is executed: Conformation Dominance Function (A... tThe calculation of the PCDPO loss function: numerically quantifies whether the immediate contribution of each decision (i.e., denoising operation) to the long-term goal during conformational evolution is "favorable" or "unfavorable" (even if the denoising gradient direction of multi-frame protein conformations is as close as possible to the conformation set rationality and the backpropagation gradient of the conformation value network). The calculation of the PCDPO loss function: based on the conformational advantage value, the optimization of the network is driven by two original loss functions: the pruning target loss function. This is the fundamental driving force behind the optimization of the policy network (i.e., the protein conformation prediction model). It employs a pruning optimization objective specifically designed for the high-dimensional continuous protein conformation space. This mechanism amplifies the impact of "favorable" decisions and suppresses "unfavorable" decisions (by dynamically calculating the diffusion denoising probability ratio r). t (θ) is used to guide the policy network to evolve towards generating better conformations. The pruning design and KL divergence term are a dual-protection mechanism specifically integrated to ensure stable exploration on complex energy surfaces while preventing the pruned target from forgetting fundamental physics knowledge. Auxiliary target loss function The objective of this auxiliary objective loss function is to train the conformational value network V. φ This allows it to evolve into a precise "physical value evaluator," capable of predicting the long-term value of any conformational state with increasing accuracy. Dual-network collaborative update: After calculating the above loss, the reinforcement learning fine-tuning algorithm simultaneously updates the policy network π via gradient descent. θ and the V-shaped value network φ The network parameters are used to complete a co-evolution.

[0254] Step 3: Target Baseline Synchronization: After the internal optimization loop is complete, the capabilities of both the policy network and the construct value network have been improved. At this point, the network parameters of the "target network" are synchronized to the latest and strongest version. This "baseline synchronization" operation prepares for the next round of more efficient "exploratory sampling".

[0255] Step 4: Experience Data Update: To ensure that the protein conformation prediction model always learns based on its latest and most cutting-edge "knowledge," it is designed as an on-policy optimization algorithm. This means that after a complete iteration of optimization, the old experience data will be cleared to ensure that the next iteration learns entirely new experiences created by a stronger version of itself.

[0256] III. Termination and Output: The main loop continues until the protein conformation prediction model's performance converges. Finally, a final policy network (i.e., the trained protein conformation prediction model) fully optimized by the PCDPO algorithm and enhanced with physical realism is output. This represents the ultimate goal: a high-level computational protein conformation prediction model that can efficiently analyze and predict protein dynamics, and whose prediction results are highly consistent with physical, chemical, and biological principles at a deep level.

[0257] The aforementioned protein conformational change prediction method based on deep generative models and reinforcement learning constructs an end-to-end, efficient deep learning proxy model to directly analyze and learn the complete dynamic process of protein conformational changes from data obtained from MD simulations. It boasts high computational efficiency and the ability to perform conformational dynamics analysis and inference at arbitrary time scales and units, collaboratively resolving the inherent contradiction between computational efficiency, physical properties, and the temporality of dynamics. Furthermore, an innovative algorithm is designed for reinforcement learning fine-tuning of the proposed protein conformational prediction model, enabling the predicted protein conformational changes (torsion angles) to better conform to dynamic behavior and physicochemical constraints without using additional training data. Simultaneously, the predicted protein conformational change process should be as close as possible to the original MD data and can be directly applied to conformational dynamics modeling of all proteins after training and reinforcement learning fine-tuning.

[0258] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages does not have to be sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0259] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0260] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting protein conformational changes based on deep generative models and reinforcement learning, characterized in that, The method includes: The L-frame random noise, first condition guidance information, and second condition guidance information are obtained. The first condition guidance information includes protein residue sequence, protein spatial contact map, and protein evolution specificity scoring matrix. The second condition guidance information is the K-frame protein initial conformation. The L-frame random noise, the first conditional guidance information, and the K-frame initial protein conformation are input into the trained protein conformation prediction model for prediction, and the L-frame protein conformation change trajectory is obtained. The training method for the protein conformation prediction model is as follows: Molecular dynamics simulation trajectory data of different proteins with 77 or less residues within 10 ns under physiological conditions and molecular dynamics simulation trajectory data of different proteins with 77 or more residues to 200 residues were obtained and preprocessed to obtain multiple sample data to form a sample dataset, each sample data including L frames. The molecular dynamics simulation trajectory data of each sample data in the sample dataset from frame K+1 to frame L are masked to obtain the masked sample dataset, which includes the masked sample data corresponding to each sample data. The sample dataset, the masked sample dataset, and the first condition guidance information are input into the protein conformation prediction model to be trained to obtain the trained protein conformation prediction model. L frames of Gaussian noise are randomly sampled, and the L frames of Gaussian noise, the first conditional guidance information and the third conditional guidance information are input into the trained protein conformation prediction model to predict the conformational change trajectory, and the protein conformational change trajectory prediction result is obtained. Based on the predicted protein conformational change trajectory, a reinforcement learning fine-tuning algorithm optimized by a protein conformational dynamics strategy is used to fine-tune the network parameters of the trained protein conformational prediction model to obtain a well-trained protein conformational prediction model.

2. The protein conformational change prediction method based on deep generative models and reinforcement learning according to claim 1, characterized in that, The protein conformation prediction model includes a multimodal local spatial perception module, a noise prediction module, and a noise reduction module. The step of inputting the L-frame random noise, the first conditional guidance information, and the K-frame initial protein conformation into a trained protein conformation prediction model for prediction to obtain the L-frame protein conformation change trajectory includes: The protein residue sequence, the protein spatial contact map, and the protein evolution-specific scoring matrix are input into the multimodal local spatial perception module for feature fusion and enhancement to obtain the fused protein dynamic features. The fused protein dynamic features, the L-frame random noise, and the K-frame initial protein conformation are input into the noise prediction module, so that the noise prediction module obtains the noise prediction result under the guidance of the fused protein dynamic features and the K-frame initial protein conformation. The random noise in the L-frame is denoised based on the noise prediction results to obtain the protein conformation change trajectory in the L-frame.

3. The protein conformational change prediction method based on deep generative models and reinforcement learning according to claim 2, characterized in that, The step of inputting the protein residue sequence, the protein spatial contact map, and the protein evolution-specific scoring matrix into the multimodal local spatial perception module for feature fusion and enhancement to obtain fused protein dynamic features includes: The protein residue sequence is input into the large prediction model of the multimodal local spatial perception module for feature extraction to obtain protein sequence features; The protein sequence features and the protein spatial contact map are fused by the feature fusion unit of the multimodal local spatial perception module, and then self-attention calculation is performed by the self-attention unit of the multimodal local spatial perception module to obtain spatially aligned sequence features. The spatially aligned sequence features and the protein evolution-specific scoring matrix are input into the cross-attention unit in the multimodal local spatial perception module for information fusion to obtain the fused protein dynamic features.

4. The protein conformational change prediction method based on deep generative models and reinforcement learning according to claim 2, characterized in that, The noise prediction module includes: a data information fusion layer, a first persistent spatiotemporal attention layer, a second persistent spatiotemporal attention layer, a third persistent spatiotemporal attention layer, and a fourth persistent spatiotemporal attention layer; The step of inputting the fused protein dynamic features, the L-frame random noise, and the K-frame initial protein conformation into the noise prediction module, so that the noise prediction module obtains noise prediction results under the guidance of the fused protein dynamic features and the K-frame initial protein conformation, includes: The L-frame random noise and the K-frame initial protein conformation are input into the data information fusion layer for fusion processing to obtain the fused protein conformation information. The fused protein dynamic features and the fused protein conformation information are input into the first persistent spatiotemporal attention layer for processing to obtain the first protein dynamic spatiotemporal features. The first protein dynamic spatiotemporal feature and the fused protein dynamic feature are input into the second persistent spatiotemporal attention layer for processing to obtain the second protein dynamic spatiotemporal feature. The second protein dynamic spatiotemporal feature and the fused protein dynamic feature are input into the third persistent spatiotemporal attention layer for processing to obtain the third protein dynamic spatiotemporal feature. The third protein dynamic spatiotemporal feature and the fused protein dynamic feature are input into the fourth persistent spatiotemporal attention layer for processing to obtain noise prediction results.

5. The protein conformational change prediction method based on deep generative models and reinforcement learning according to claim 4, characterized in that, The first persistent spatiotemporal attention layer, the second persistent spatiotemporal attention layer, the third persistent spatiotemporal attention layer, and the fourth persistent spatiotemporal attention layer are persistent spatiotemporal attention networks with the same structure. The persistent spatiotemporal attention network includes spatial cross attention, temporal self-attention, spatial self-attention, and a feedforward neural network connected in sequence.

6. The protein conformational change prediction method based on deep generative models and reinforcement learning according to claim 1, characterized in that, The training objective of the protein conformation prediction model to be trained is: in, The loss function during the training phase. For expectation, ∈ α This represents the actual noise distribution during the noise addition process. For the sample dataset, x t Let L be the protein conformation of the frame to be denoised at time step t, where t represents a continuous uniform distribution from 0 to 1. The time step for uniform sampling is defined by c_guide, which represents the input guiding condition information. θ represents the predicted noise distribution, and θ represents the network parameters of the protein conformation prediction model.

7. The protein conformational change prediction method based on deep generative models and reinforcement learning according to claim 1, characterized in that, The reinforcement learning fine-tuning algorithm for optimizing the protein conformation dynamics strategy includes: Based on energy reward, conflict detection reward, Ramachandran reward, secondary structure preservation reward, and conformational diversity reward, a comprehensive reward function incorporating physical, geometric, and biological constraints is constructed. Based on the probability ratio of single-step conformation transitions in the trained protein conformation prediction model, a pruning target loss function is constructed. Based on the comprehensive reward function, an auxiliary target loss function is constructed using the constructive value network; Based on the pruning target loss function and the auxiliary target loss function, construct the overall optimization target function; The network parameters of the trained protein conformation prediction model and the network parameters of the conformation value network are updated collaboratively according to the overall optimization objective function until the performance of the trained protein conformation prediction model converges, thus obtaining a well-trained protein conformation prediction model.

8. The protein conformational change prediction method based on deep generative models and reinforcement learning according to claim 7, characterized in that, The overall optimization objective function for: in, To prune the target loss function, λ v The weighting coefficients of the loss term in the auxiliary objective loss function, To assist the target loss function, λ KL Let be the weight coefficient of a KL divergence regularization term. Here, φ is a KL divergence regularization term, φ is the network parameter of the conformational value network, and θ is the network parameter of the protein conformation prediction model.

9. The protein conformational change prediction method based on deep generative models and reinforcement learning according to claim 8, characterized in that, The auxiliary target loss function is: Among them, V φ (x T (c_guide) represents the value reward score for the conformational value network assessment, x T L frames of Gaussian noise, c_guide is the input guidance condition information, R(x0) is the comprehensive reward function, and proein~S is the protein sequence obtained by randomly sampling from the protein sequence library S with different residue sequences.

10. The protein conformational change prediction method based on deep generative models and reinforcement learning according to claim 9, characterized in that, The target loss function for pruning is: Where T is the maximum time step, r t (θ) represents the probability ratio of single-step protein conformational transfer, A t Let clip_b be the conformational dominance function at time step t, and clip_b be a hyperparameter defining the clipping boundary; clip(·) represents clipping.

Images