A method and system for protein structure generation and optimization by diffusion

Abstract

The application relates to the technical field of biological information, in particular to a protein structure diffusion generation and optimization method and system, which comprises the following steps: obtaining an amino acid sequence of a protein and converting and extracting sequence features; fusing sequence feature representation and a spherical harmonic function to a convolutional neural network to construct a convolution module; performing a noise adding and unfolding operation on a protein conformation, executing a forward prediction process, and outputting a design latent code; performing a latent code space operation on the design latent code and performing a reverse denoising and folding operation, executing a reverse generation process, and outputting a predicted conformation; performing bidirectional correction on the forward prediction process and the reverse generation process, constructing a diffusion generation model of a protein structure, and obtaining a final three-dimensional structure result of the protein. Through the application, the problems of insufficient utilization of multi-time step information and lack of coordination of forward and reverse evolution in protein structure diffusion generation are effectively solved, the consistency of structure evolution trajectory is realized, and the generation stability and prediction accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method and system for the diffusion generation and optimization of protein structures. Background Technology

[0002] The three-dimensional structure of proteins is directly related to their biological functions. Therefore, modeling and optimizing protein three-dimensional structures using computational methods has become an important research direction in structural biology and artificial intelligence. In recent years, diffusion models have been introduced into protein structure generation tasks, showing great potential in protein structure sampling.

[0003] Most existing diffusion sampling methods are based on single-step Euler methods or simple numerical integration schemes, such as Generative Diffusion Models (RGDM) based on inverse renormalized group flow and Denoising Diffusion Probability Models (DDPM). However, these sampling methods have slow convergence speeds, require a large number of iterations, and lack accuracy in predicting complex protein topologies, making it difficult to meet the requirements for high-precision structure prediction. Furthermore, existing methods cannot simultaneously optimize local secondary structure details and global folding patterns. RGDM has weak ability to capture high-frequency details, and DDPM lacks protein-specific physical constraints, resulting in poor multi-scale feature fusion performance. How to design efficient sampling algorithms that can significantly reduce sampling steps while ensuring prediction accuracy is the core challenge facing current diffusion methods for protein structure prediction. Summary of the Invention

[0004] This invention provides a method for the diffusion generation and optimization of protein structures, which can effectively solve the problems in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for protein structure diffusion generation and optimization, the method comprising: The amino acid sequence of the protein is obtained and transformed to extract sequence features, which represent contextual semantics, structural preferences and evolutionary conservation information used to describe the amino acid sequence, including edge features and node features; The sequence feature representation is compared with the spherical harmonic function of the equivariant neural network. Integrate it into a convolutional neural network to construct a convolutional module for predicting the isotropic 3D coordinates of the current protein; The Adams-Bashforth multi-step method is applied to the protein conformation to perform a noisy defolding operation, a forward prediction process is executed, and the design latent code is output. Perform latent code space operations on the designed latent code, apply the Adams-Moulton multi-step method for inverse denoising and folding operations, perform inverse generation process, and output predicted conformation; A bidirectional explicit linear multistep method is used to bidirectionally correct the forward prediction process and the reverse generation process to construct a diffusion generation model for protein structures. Based on the convolution module and the diffusion generation model, the final three-dimensional structure of the protein is obtained.

[0006] Furthermore, the sequence feature representation is compared with the spherical harmonic function of the equivariant neural network. Integrating into a convolutional neural network, a convolutional module is constructed to predict the isotropic 3D coordinates of the current protein, including: The sequence feature representation is subjected to feature mapping processing to obtain a high-dimensional feature representation for convolution operations; Based on the relative spatial relationships between protein residues, the corresponding spatial geometric information is calculated, and a spherical harmonic function representation is generated based on the spatial geometric information. The high-dimensional feature representation is fused with the spherical harmonic function representation to obtain a fused feature representation that satisfies spatial rotation equivariance and translation invariance. Based on the fusion feature representation, perform equivariant convolution operation to extract spatial features related to the current three-dimensional structural state of the protein while maintaining the equivariance of three-dimensional space. Based on the result of the isovariant convolution operation, the isovariant three-dimensional coordinate update amount corresponding to the current three-dimensional structural state of the protein is predicted.

[0007] Furthermore, the convolutional neural network includes: An embedding layer is used to receive the protein structural feature representation and map the protein structural feature representation to a high-dimensional feature space; The geometric feature processing layer is used to extract spatial geometric features and calculate spherical harmonic function representations based on the relative positional relationships between residues in a three-dimensional spatial coordinate system. The equivariant convolutional layer performs joint feature extraction on the protein structure feature representation and the spatial geometric features through equivariant convolution and tensor product operations, while maintaining the equivariance of features under spatial rotation and translation transformations. The output layer is used to predict the structural state update vector corresponding to the three-dimensional structural state of the protein based on the feature representation output by the isovariant convolutional layer.

[0008] Furthermore, after performing the forward explicit linear multistep prediction update, the method also includes attribute interpolation, directional editing, and condition constraint processing on the structural latent code representation.

[0009] Furthermore, the forward explicit linear multi-step prediction update includes: Obtain the three-dimensional structural state at multiple historical time steps and its corresponding structural state update vector; The predicted update amount for the current time step is obtained by linearly combining the structure state update vectors of the multiple historical time steps. The predicted update amount is numerically synthesized with the three-dimensional structural state of the most recent time step, and the three-dimensional structural state of the most recent time step is updated to obtain the structural latent code representation corresponding to the current time step.

[0010] Furthermore, the reverse explicit linear multi-step correction update includes: The structural latent code is used as the starting state for reverse derivation; On the timeline, in order of time steps earlier than the current time step, obtain the three-dimensional structural state and the corresponding structural state update vector for multiple reverse historical time steps; Based on the three-dimensional structural states of the multiple reverse historical time steps and the corresponding structural state update vectors, the three-dimensional structural state of the protein in the current reverse time step is corrected and updated. The three-dimensional structure state is gradually restored from back to front by sequentially traversing each reverse time step along the time axis and performing the correction update at each reverse time step.

[0011] Furthermore, a diffusion generation model for protein structures is constructed by bidirectionally coupling the forward explicit linear multistep prediction results with the reverse explicit linear multistep correction update results, including: According to the preset bidirectional fusion weights, the forward explicit linear multistep prediction update result and the reverse explicit linear multistep correction update result are weighted and fused to obtain the coupled structure state at the corresponding time step. The weighted fusion is performed sequentially along the time step sequence to obtain a bidirectional coupled structural state sequence covering the entire diffusion generation process; Based on the bidirectional coupled structural state sequence, model parameters for describing the evolution of the three-dimensional structural state of the protein are determined, and the diffusion generation model is obtained based on the model parameters.

[0012] Furthermore, when bidirectionally coupling the forward explicit linear multistep prediction update result and the reverse explicit linear multistep correction update result, the method further includes assigning adjustable fusion weights to both, and adaptively adjusting the fusion weights according to the time step position of the three-dimensional structure state.

[0013] A protein structure diffusion generation and optimization system, the system comprising: The feature representation construction module obtains the amino acid sequence of a protein and transforms it to extract sequence features. The sequence feature representation is used to describe the contextual semantics, structural preferences and evolutionary conservation information of the amino acid sequence, including edge features and node features. The updated vector prediction module combines the sequence feature representation with the spherical harmonic function of the equivariant neural network. Integrate it into a convolutional neural network to construct a convolutional module for predicting the isotropic 3D coordinates of the current protein; The forward prediction update module applies the Adams-Bashforth multi-step method to perform noisy defolding operations on the protein conformation, executes the forward prediction process, and outputs the design latent code. The reverse correction update module performs latent code space operations on the designed latent code, applies the Adams-Moulton multi-step method to perform reverse denoising and folding operations, performs the reverse generation process, and outputs the predicted conformation. The bidirectional result coupling module uses a bidirectional explicit linear multi-step method to perform bidirectional correction on the forward prediction process and the reverse generation process, thereby constructing a diffusion generation model for protein structures. The structure result generation module, based on the convolution module and the diffusion generation model, obtains the final three-dimensional structure result of the protein.

[0014] Furthermore, the bidirectional result coupling module includes: The result weighted fusion unit performs weighted fusion of the forward explicit linear multistep prediction update result and the reverse explicit linear multistep correction update result according to the preset bidirectional fusion weights to obtain the coupled structure state at the corresponding time step. The state sequence acquisition unit performs the weighted fusion sequentially along the time step sequence to obtain the bidirectional coupled structure state sequence of the entire coverage diffusion generation process; The diffusion model generation unit determines model parameters to describe the evolution of the three-dimensional structural state of the protein based on the bidirectional coupled structural state sequence, and obtains the diffusion generation model based on the model parameters.

[0015] The technical solution of this invention can achieve the following technical effects: This method effectively addresses the problems of insufficient utilization of structural states at multiple historical time steps in the existing protein structure diffusion generation process, as well as the lack of time trajectory coordination constraints between the forward generation and reverse recovery processes. By introducing an explicit linear multi-step update method during structural evolution and combining bidirectional coupled modeling of forward structural extrapolation and reverse structural reconstruction, the evolutionary relationship of the protein's three-dimensional structural state over time can be more accurately characterized, thereby significantly improving the stability and accuracy of structure generation during the diffusion generation process.

[0016] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for protein structure diffusion generation and optimization. Figure 2 This is a schematic diagram of the diffusion model; Figure 3 A flowchart illustrating the process of constructing a convolutional module for predicting the isotropic 3D coordinates of the current protein; Figure 4 This is a cross-sectional view of a convolutional neural network. Figure 5 This is a schematic diagram of the forward explicit linear multi-step prediction update process; Figure 6 A schematic diagram of the reverse explicit linear multi-step correction update process; Figure 7 A schematic diagram of the process for constructing a diffusion generation model for protein structures. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Example 1: like Figure 1 As shown, this application provides a method for protein structure diffusion generation and optimization, the method comprising: S1: Obtain the amino acid sequence of the protein and extract sequence features. Sequence features represent contextual semantics, structural preferences, and evolutionary conservation information used to describe the amino acid sequence, including edge features and node features. Specifically, after obtaining the amino acid sequence of the protein, the amino acid sequence undergoes standardized preprocessing, including encoding amino acid characters, handling abnormal or missing sites, and preserving the original order information of amino acids in the sequence to ensure that the sequence feature construction process can fully reflect the primary structure information of the protein. After sequence preprocessing, feature extraction is performed on the amino acid sequence, mapping each amino acid residue to a corresponding node feature. The node feature includes at least feature information reflecting the physicochemical properties of the amino acid, contextual semantic features, and preference features related to structure formation, enabling the node feature to characterize the functional role and structural tendency of a single residue in the sequence environment. Simultaneously, edge features are constructed based on the relative positional relationships and potential interactions of amino acids in the sequence. Edge features are used to describe the association information between different amino acid residues, including at least sequence distance information, adjacency information, and feature parameters reflecting the strength of mutual influence between residues, thereby establishing an association representation between residues at the sequence level. During the sequence feature extraction process, information reflecting the evolutionary conservation of amino acids is introduced. By encoding the degree of conservation of amino acids during evolution, the sequence features can reflect the evolutionary constraints of proteins in different species or different sequence backgrounds, thereby enhancing the ability of sequence features to express the laws of structure formation. Through the above processing, a sequence feature representation that simultaneously contains node features and edge features is constructed. This allows the sequence feature representation to comprehensively characterize the contextual semantics, structural preferences, and evolutionary conservation information of amino acid sequences without relying on three-dimensional structural information, and to serve as the basic input for subsequent three-dimensional structural modeling and diffusion generation processes.

[0022] S2: Representing sequence features with the spherical harmonic function of an equivariant neural network Integrate it into a convolutional neural network to construct a convolutional module for predicting the isotropic 3D coordinates of the current protein; Specifically, after obtaining the sequence feature representation constructed in step S1, an isovariant neural network is introduced to model the three-dimensional spatial geometric relationship of the protein. By inputting the sequence feature representation and the spherical harmonic function representation constructed based on three-dimensional spatial orientation information into the convolutional neural network, the network considers both sequence semantic information and spatial geometric information simultaneously during feature learning. Specifically, based on the relative positional relationship of each residue in the protein in three-dimensional space, geometric information reflecting spatial orientation and angular distribution is extracted and encoded into the form of spherical harmonic functions. This allows spatial orientation features to participate in convolution calculation in a mathematically describable manner, thereby providing the network with explicit three-dimensional spatial structural constraints. In the convolutional neural network, the sequence feature representation and the spherical harmonic function representation are jointly processed, fusing sequence context information and spatial orientation information in the same feature calculation process. The network is equipped with information that enables it to perceive the spatial arrangement of residues during feature propagation and aggregation, and to make corresponding isotropic responses to feature changes when rotational transformations occur in three-dimensional space, while maintaining feature expression invariance under translational transformations. By performing isotropic convolution operations on the fused features, the convolution calculation results can accurately reflect the spatial correlation characteristics of the current three-dimensional conformation of the protein, and thus learn the trend of protein structure changes in three-dimensional space without relying on an absolute coordinate reference frame. Based on the feature results obtained from the isotropic convolution operation, the network outputs isotropic three-dimensional coordinate prediction results corresponding to the current three-dimensional structural state of the protein, ensuring that the predicted three-dimensional coordinates maintain an isotropic relationship under spatial rotation conditions, and serving as the basic input for the three-dimensional structure evolution in subsequent diffusion generation, structure update or correction steps.

[0023] S3: Apply the Adams-Bashforth multi-step method to the protein conformation to perform noisy defolding operations, execute the forward prediction process, and output the designed latent code; Specifically, the current three-dimensional structural state corresponding to the protein conformation is used as the initial state for forward evolution. Gradually increasing random perturbations are introduced along the diffusion time axis to add noise to the protein conformation, causing it to evolve from its native folded conformation to a disordered or semi-disordered state. This provides a continuous and controllable diffusion trajectory for the structure generation process. During forward prediction, the Adams-Bashforth multi-step method is introduced to numerically update the protein conformation evolution. The Adams-Bashforth multi-step method explicitly extrapolates the structural change trend of the current time step by comprehensively utilizing structural state change information from multiple historical time steps. This ensures that conformational evolution no longer relies solely on local information from a single time step but reflects the continuous change pattern of protein conformation over time. Specifically, in the initial stage of forward prediction, low-order prediction is used to gradually accumulate historical time step information based on the existing number of historical conformations. Once the number of historical time steps meets the multi-step prediction condition, a multi-step Adams-Bashforth method is employed. The prediction formula linearly combines the structural change directions corresponding to multiple historical time steps to calculate the conformational update amount at the current time step, thereby achieving forward prediction and update of the protein conformation. During the multi-step prediction and update process, the calculated conformational update amount is numerically synthesized with the current protein conformation, so that the protein conformation gradually deviates from the natural folded state while maintaining continuity, achieving unfolding and noise addition effects. By controlling the diffusion time step and noise intensity, the conformational evolution process smoothly transitions from order to disorder, avoiding numerical oscillations or non-physical mutations. After multiple forward prediction and updates, the protein conformation is mapped to the diffusion latent space, forming a design latent code to represent the protein structural evolution state. The design latent code retains the overall topological features of the original conformation while weakening local geometric constraints, making it suitable as the input representation for subsequent latent code space operations and reverse generation processes.

[0024] S4: Perform latent code space operations on the designed latent code, apply the Adams-Moulton multi-step method for inverse denoising and folding operations, perform the inverse generation process, and output the predicted conformation; Specifically, the design latent code obtained in step S3 is used as the initial representation of the reverse generation process. Spatial operations are performed on the design latent code within the latent code space to allow controlled changes in the latent code while maintaining overall structural consistency, thereby providing a target-oriented potential structural state for subsequent reverse generation. During the latent code space operation, attribute interpolation, directional editing, or constraint projection operations are performed on the design latent code according to the preset generation target or structural constraints. By linearly or nonlinearly adjusting the latent code vector, the latent code is shifted along a specific direction in the latent space, thus affecting the structural feature distribution of the subsequently generated conformation. Through latent code space operations, the overall morphology, local conformational features, or functionally related attributes of the generated structure are guided without disrupting the continuity and generativeness of the latent code representation. After completing the latent code space operation, the adjusted design latent code is used as the starting state for reverse evolution. A reverse generation process from disorder to order is performed on the diffusion time axis, and the Adams-Moulton multi-step method is introduced to numerically update the reverse evolution of the protein conformation. The Adams-Moulton multistep method comprehensively utilizes conformational change information from multiple backward historical time steps to correct and update the conformational state at the current time step. This allows the backward generation process to not only rely on the current prediction results but also to incorporate the overall evolutionary trajectory for correction. In the initial stage of backward generation, low-order correction methods are used to gradually accumulate backward time step information based on the number of available backward historical conformations. Once the number of backward historical time steps meets the multistep correction condition, the multistep Adams-Moulton correction formula is used to jointly calculate the conformational change trends corresponding to multiple backward time steps, obtaining the conformational correction amount for the current time step. This gradually maps the conformation from the latent space back to a stable three-dimensional structural state. During the backward denoising and folding process, the conformational correction amount is numerically synthesized with the current conformational state, allowing the protein conformation to gradually recover its folded structure while maintaining continuous evolution, while suppressing the numerical errors and random perturbations accumulated during forward prediction. By controlling the backward time step size and correction intensity, the conformational folding process is made smooth and stable, avoiding structural collapse or non-physical conformations. After multiple reverse correction updates, a predicted conformation corresponding to the designed latent code is obtained. The predicted conformation is consistent with the physical folding law of the protein in terms of overall topology and local geometric features, and serves as the input result for subsequent bidirectional coupling or structural evaluation processes.

[0025] S5: A diffusion generation model for protein structures is constructed by using a bidirectional explicit linear multi-step method to perform bidirectional correction on the forward prediction process and the reverse generation process. Specifically, Figure 2The schematic diagram of the diffusion model treats two temporal evolution trajectories as complementary and mutually constraining sources of structural representation. During model training and generation, forward prediction sequences and backward recovery sequences are introduced simultaneously. A bidirectional fusion mechanism establishes a mapping relationship between the two, enabling the model to learn the entire behavioral pattern of structural transition from ordered conformations to latent code space and back to stable conformations within a unified framework. In a preferred implementation, based on the time-step-level structural correspondence, the structural latent code obtained from forward prediction and the reconstructed structure obtained from backward correction are synchronously fused. This ensures that structural change information at different time scales is consistently expressed within the model, thus forming a diffusion-generative model with both extrapolation and backtracking capabilities.

[0026] S6: Based on the convolution module and the diffusion generation model, the final three-dimensional structure of the protein is obtained.

[0027] Specifically, after completing forward prediction, backward generation, and bidirectional coupling modeling, the isovariant three-dimensional coordinate update information output by the convolution module and the structural evolution results obtained by the diffusion generation model are integrated and processed to finally determine the three-dimensional conformation of the protein, so that the generation results simultaneously satisfy sequence constraints, spatial geometric constraints, and diffusion evolution laws. Specifically, based on the isovariant 3D coordinate updates predicted by the convolution module at each time step, the protein conformation is progressively corrected, ensuring that the positional changes of residues in 3D space maintain rotational isovariance and translational invariance, thereby avoiding dependence of the generated result on a specific coordinate reference frame. By continuously accumulating coordinate updates, the overall protein conformation forms a stable and coherent 3D structural layout in space. Simultaneously, by combining the forward prediction trajectory and backward correction trajectory of the diffusion generation model on the time axis, the evolution process of the protein conformation is constrained as a whole, ensuring that the final conformation conforms to the structural change trend learned during forward diffusion and satisfies the stable folding constraints formed during backward generation, thereby improving the rationality of the generated structure at the global topological and local geometric levels. When determining the final 3D structure result, consistency checks and numerical stability processing can be performed on the generated conformation, including constraining or adjusting the distance distribution between residues, local geometric continuity, and overall conformation smoothness, to eliminate minor deviations that may be introduced during numerical evolution, making the generated structure more consistent with the physical folding characteristics of proteins. Through the above processing, the final three-dimensional structure of the protein is output. The generated three-dimensional structure can accurately reflect the relative positional relationship of residues in three-dimensional space while maintaining the stability of the overall folding pattern of the protein. It can also serve as the final output for protein structure prediction, structure optimization, or subsequent functional analysis.

[0028] This invention effectively solves the problems of insufficient utilization of multi-time step information and lack of coordination between forward and reverse evolution in protein structure diffusion generation, achieves consistency of structure evolution trajectory, and improves generation stability and prediction accuracy.

[0029] As a preferred embodiment of the above, such as Figure 3 As shown, in step S2, the sequence feature representation is compared with the spherical harmonic function of the equivariant neural network. Integrating into a convolutional neural network, a convolutional module is constructed to predict the isotropic 3D coordinates of the current protein, including: S21: Perform feature mapping on the sequence feature representation to obtain a high-dimensional feature representation for convolution operations; S22: Based on the relative spatial relationships between protein residues, calculate the corresponding spatial geometric information and generate a spherical harmonic function representation based on the spatial geometric information; S23: The high-dimensional feature representation and the spherical harmonic function representation are fused to obtain a fused feature representation that satisfies spatial rotation equivariance and translation invariance; S24: Based on the fusion feature representation, perform equivariant convolution operation to extract spatial features related to the current three-dimensional structural state of the protein while maintaining the equivariance of three-dimensional space; S25: Based on the result of the isovariant convolution operation, predict the isovariant three-dimensional coordinate update amount corresponding to the current three-dimensional structural state of the protein.

[0030] Specifically, feature mapping is performed on the sequence feature representation. This involves taking the sequence features corresponding to the protein amino acid sequence as input and using a combination of linear mapping and nonlinear activation to map the original sequence features from a low-dimensional or non-uniform expression space to a unified high-dimensional feature space. This makes sequence information from different sources and scales more suitable for subsequent convolution calculations in terms of numerical distribution and representation, thus providing a stable and continuous feature foundation for 3D geometric feature modeling. Spatial geometric information is calculated based on the relative spatial relationships between protein residues, where the relative spatial relationships are determined by the coordinate differences of the residues in 3D space. The spatial geometric information includes at least the relative displacement vectors between residues, the distance, and the direction and angle parameters determined by the relative displacement vectors. After obtaining the spatial geometric information, it is encoded according to the 3D spatial orientation characteristics to generate a spherical harmonic function representation that can characterize the spatial orientation distribution features. This introduces the spatial geometric relationships into the feature modeling process in a mathematical form with clear physical meaning. In step S23, the high-dimensional feature representation obtained in step S21 is compared with that obtained in step S22. The obtained spherical harmonic function representations are fused. This fusion process involves jointly representing sequence features and spatial orientation features within a unified feature space. This ensures that the fused features exhibit an isotropic response during rotational transformations in three-dimensional space and remain numerically unchanged during translational transformations. This simultaneously preserves sequence semantic information and spatial geometric information, ensuring that the feature expression remains consistent with the actual spatial conformational changes of the protein. (Step S24) In this process, equivariant convolution operations are performed based on fused feature representations. The equivariant convolution operation explicitly constrains the feature transformation relationship during feature propagation and neighborhood aggregation, ensuring that the convolution calculation results meet the equivariant requirements under three-dimensional spatial rotation and translation transformations. By performing convolution calculations on the fused features within each residue and its neighborhood in the protein, spatial correlation features reflecting the current three-dimensional structural conformation of the protein are extracted. These spatial correlation features can describe the relative positional changes between residues and the overall conformational evolution trend. Finally, based on the spatial correlation features obtained from the equivariant convolution operation, the equivariant three-dimensional coordinate update amount corresponding to the current three-dimensional structural state of the protein is predicted. The equivariant three-dimensional coordinate update amount is used to characterize the displacement direction and displacement amplitude of each residue in the protein in three-dimensional space, ensuring that the coordinate update maintains a consistent equivariant relationship under spatial rotation transformations, and serving as the basis for updating the three-dimensional structural state of the protein in subsequent diffusion generation, forward prediction, or backward correction processes.

[0031] As a preferred embodiment of the above, the convolutional neural network includes: The embedding layer is used to receive protein structural feature representations and map these representations to a high-dimensional feature space. The geometric feature processing layer is used to extract spatial geometric features and calculate spherical harmonic function representations based on the relative positional relationships between residues in a three-dimensional spatial coordinate system. Equivariant convolutional layers extract joint features from protein structural features and spatial geometric features through equivariant convolution and tensor product operations, while maintaining the equivariance of features under spatial rotation and translation transformations. The output layer is used to predict the structural state update vector corresponding to the three-dimensional structural state of the protein based on the feature representation output by the isovariant convolutional layer.

[0032] Specifically, Figure 4 This is a cross-sectional view of a convolutional neural network. The embedding layer receives protein structural feature representations and maps the input residue sequence information, physicochemical properties, and spatial location information to a high-dimensional feature space. This encodes structural information from different sources into an internal representation with a consistent vector format, providing a foundational expressive capability for subsequent deep structure modeling. Subsequently, the geometric feature processing layer analyzes the relative positional relationships between residues in a three-dimensional coordinate system. By calculating geometric quantities such as distance, direction, and angle, it forms spatial geometric features and generates structural representations in the form of spherical harmonic functions based on these geometric quantities. This allows the network to explicitly focus on three-dimensional geometric structures, rather than just sequence relationships. After completing the geometric feature construction, isovariant convolutional layers process the embedded structural features and spatial geometric features... Joint convolutional computation is performed, and this convolutional layer preferably employs a convolutional framework with rotational and translational equivariance properties. This ensures that structural features maintain a stable and consistent response even when rotated or translated in the spatial coordinate system, thereby ensuring that the structural rules learned by the network are only related to the true conformation and do not depend on a specific coordinate reference frame. Furthermore, through tensor product operations, feature interaction and coupling expression are achieved between scalar features and vector features of different orders and higher-order tensor features to more fully characterize the complex three-dimensional structural relationships between residues. Finally, the deep features obtained from the equivariant convolutional layer are analyzed and mapped through the output layer, transforming the high-dimensional structural expression into a structural state update vector corresponding to the current three-dimensional structural state of the protein. This vector indicates the direction and magnitude of geometric adjustments that the three-dimensional structure should undergo during diffusion generation. This implementation introduces an equivariant neural network structure, enabling the convolutional neural network to possess physical consistency and geometric robustness in three-dimensional space, thereby more accurately modeling the spatially dependent features in the protein structure evolution process.

[0033] As a preferred embodiment of the above, after performing forward explicit linear multistep prediction update, the method further includes attribute interpolation, directional editing, and condition constraint processing on the structural latent code representation.

[0034] Specifically, attribute interpolation can weightedly combine two or more structural latent code samples from different sources, enabling a continuous transition between different conformations or functional attributes in the generated results. This results in novel structural forms with intermediate attribute features, such as achieving a balance between indicators like stability, flexibility, or hydrophilicity. Directional editing introduces latent directions related to specific structural or functional attributes into the latent code space, directionally shifting the structural latent code so that the generated structure shifts along the target attribute direction. This allows for targeted optimization of key structural features while maintaining the overall conformational framework. Furthermore, it can also be performed within the latent code space... By implementing conditional constraints, the generation space is limited to a subspace that meets preset geometric, functional, or biophysical constraints. For example, the proportion of secondary structures, distance distribution patterns, or conformational stability at specific sites are restricted to avoid latent code shifts that cause the structure to deviate from the true foldable conformation region. Through the aforementioned latent code space operation mechanism, this invention can not only achieve the natural diffusion generation of protein structures, but also introduce functional guidance or design goals during the generation stage. This makes the final output structure not only physically reasonable but also more in line with application requirements, thereby further expanding the applicability of the entire method in protein functional design, structure optimization, and new molecule construction.

[0035] As a preferred embodiment of the above, such as Figure 5 As shown, the forward explicit linear multi-step prediction update includes: A10: Obtain the three-dimensional structural state and its corresponding structural state update vector at multiple historical time steps; A20: Based on the linear combination calculation of the structural state update vectors of multiple historical time steps, the predicted update amount for the current time step is obtained; A30: The predicted update amount is numerically synthesized with the three-dimensional structural state of the most recent time step, and the three-dimensional structural state of the most recent time step is updated to obtain the structural latent code representation corresponding to the current time step.

[0036] Specifically, firstly, the corresponding three-dimensional structural states and the structural state update vectors predicted by the convolutional neural network at multiple historical time steps that have already been extrapolated are obtained. These historical time steps are continuously distributed before the current time point and can collectively reflect the natural evolution trend of protein structure in the time dimension. Subsequently, based on the structural state update vectors of multiple historical time steps, these update vectors are numerically weighted and combined through a preset or learned linear combination strategy, so that the update contribution of each historical time step is reasonably reflected, thereby forming the predicted update amount required for the current time step. In this way, the current structural change not only depends on the information of the latest time step, but also makes full use of the structural evolution trajectory accumulated in earlier time steps, making the prediction results smoother, more continuous and stable in the time dimension. Further, the predicted update amount and the three-dimensional structural state of the latest time step are numerically synthesized to update the structural state of the most recent time step, so that the structure completes the extrapolation evolution along the prediction direction at the current time step and generates the corresponding structural latent code representation, which serves as the time node structural input in the subsequent generation process.

[0037] As a preferred embodiment of the above, such as Figure 6 As shown, the reverse explicit linear multi-step correction update includes: B10: The structural latent code is used as the starting state for reverse derivation; B20: On the timeline, obtain the three-dimensional structural states and corresponding structural state update vectors of multiple reverse historical time steps in order of time steps earlier than the current time step; B30: Based on the three-dimensional structural states of multiple reverse historical time steps and the corresponding structural state update vectors, the three-dimensional structural state of the protein in the current reverse time step is corrected and updated. B40: Iterate through each reverse time step along the time axis, perform correction updates at each reverse time step, and gradually restore the three-dimensional structure state from back to front.

[0038] Specifically, firstly, the structural latent code representation is used as the initial input to the reverse derivation process, which is considered as the structural state when the diffusion generation process reaches the target termination time step. Then, along the time axis, the three-dimensional structural states and their corresponding structural state update vectors of multiple recovered or derived reverse historical time steps are sequentially acquired in order of precedence to the current time step. These reverse historical time steps collectively reflect the evolutionary trend of the structure along the backtracking path. Next, based on the structural states and update vectors of multiple reverse historical time steps, an explicit linear multi-step correction mechanism is used to correct the three-dimensional structural state of the current reverse time step, thereby reducing the offset caused by error accumulation during the structural backtracking process. This allows the current back-reasoning result to better match the actual structural evolution trajectory. In this process, the reverse multi-step correction does not only rely on the state of a single time step but also considers the correlation between multiple reverse time nodes to enhance the stability and consistency of structural recovery. Finally, each reverse time step is traversed sequentially from back to front along the time axis, and a correction update operation is performed at each reverse time step, allowing the structure to gradually backtrack from the latent code space to a stable conformational state in the time dimension, forming a continuous and traceable reverse structural recovery path. Through the above-described reverse explicit linear multi-step correction and update process, this invention can make full use of time history information in the structural backtracking stage, effectively suppress the problem of deviation accumulation caused by single-step backtracking, and greatly improve the stability and accuracy of the structural reconstruction stage.

[0039] Furthermore, in existing technologies, the temporal evolution of structural states during protein structure diffusion generation is mostly achieved through single-step or local time-step updates. This approach has limited utilization of structural states across multiple historical time steps and fails to fully reflect the continuous evolution of protein structures over time. Meanwhile, the forward diffusion process and the reverse recovery process are typically modeled independently, lacking consistency constraints based on time trajectories between them. This leaves room for improvement in the structural convergence and stability of the generation process.

[0040] To solve the above problems, such as Figure 7 As shown, step S5 involves bidirectionally coupling the forward explicit linear multi-step prediction results with the reverse explicit linear multi-step correction update results to construct a protein structure diffusion generation model, including: S51: Based on the preset bidirectional fusion weights, the forward explicit linear multistep prediction update result and the reverse explicit linear multistep correction update result are weighted and fused to obtain the coupled structure state at the corresponding time step. S52: Perform weighted fusion sequentially along the time step sequence to obtain the bidirectional coupled structure state sequence covering the entire diffusion generation process; S53: Based on the bidirectional coupled structural state sequence, determine the model parameters used to describe the evolution relationship of the three-dimensional structural state of proteins, and obtain the diffusion generation model based on the model parameters.

[0041] Specifically, firstly, at each time step, according to a preset bidirectional fusion weight, the structural state predicted by the forward multi-step method and the structural state corrected by the backward multi-step method are weighted and fused, so that structural information from both sources participates in the construction of the structural representation at that time step. The forward result mainly reflects the extrapolation trend of the structure as it diffuses into the latent code space, while the backward result reflects the convergence law of the structure as it returns to a stable conformation from the latent code space. The combination of the two can form a coupled structural state with stronger physical interpretability and temporal consistency. Subsequently, the weighted fusion operation is performed sequentially along the time step sequence to generate the coupled structural state for each time step in the diffusion process. This approach forms a bidirectional coupled structural evolution trajectory covering the entire diffusion process, transforming diffusion generation from a unidirectional random drift into a continuous, smooth, and self-consistent structural change path under the combined constraints of extrapolation and backtracking. Finally, the entire bidirectional coupled structural state sequence is used as a time-series learning object. Through parameter learning or probabilistic modeling, the statistical characteristics and dynamic laws of protein three-dimensional structure evolution over time are fitted to determine the model parameters describing diffusion generation behavior. Based on this, a diffusion generation model is formed, enabling the model to infer and sample structural states in subsequent generation processes based on the learned parameter laws. Through this bidirectional coupled modeling method, this invention unifies forward prediction and backward backtracking within the same structural evolution framework, giving the diffusion generation process stronger reversibility, continuity, and stability, thereby significantly improving the reliability and accuracy of protein structure generation results.

[0042] As a preferred embodiment of the above, when bidirectionally coupling the forward explicit linear multistep prediction update result and the reverse explicit linear multistep correction update result, the method further includes assigning adjustable fusion weights to both respectively, and adaptively adjusting the fusion weights according to the time step position of the three-dimensional structure state.

[0043] Specifically, when bidirectionally coupling the forward explicit linear multistep prediction update results and the backward explicit linear multistep correction update results, an adjustable fusion weight mechanism can be further introduced. That is, adjustable fusion weights are assigned to the forward structure extrapolation results and the backward structure backtracking results respectively, and the weights are adaptively adjusted according to the time step position of the three-dimensional structure state, so as to dynamically balance the exploratory structure changes and the stability convergence trend throughout the diffusion generation process. In the early stages of diffusion, the 3D structural state is closer to the natural conformation or initial state, requiring the model to have strong structural exploration capabilities to fully search the potential conformation space. Therefore, the fusion weight of the forward prediction results can be increased first, allowing the coupled structural state to extrapolate and evolve more along the forward direction. In the later stages of diffusion, the 3D structural state is close to the termination state of the latent code space or nearing the completion stage of generation. It is more necessary to rely on the back-correction update results to guide the structure to gradually return to a stable conformation. At this time, the fusion weight of the back-correction results can be gradually increased to make the structural backtracking path smoother and more stable, reducing the cumulative offset caused by random perturbations. At the same time, the fusion weight can be dynamically adjusted according to the structural change rate, stability evaluation index, or convergence characteristics during the training phase. This makes the weight allocation no longer a fixed setting, but can automatically adapt to changes in time step and structural state, thus forming a bidirectional fusion mode that balances stability and flexibility. Through this adaptive fusion weight mechanism, the bidirectional coupling process can play a differentiated role at different stages, enabling the diffusion generation model to have sufficient exploration capabilities in the early stages and strong convergence capabilities in the later stages, significantly improving the consistency and physical rationality of the overall structural generation trajectory.

[0044] Example 2: Based on the same inventive concept as the protein structure diffusion generation and optimization method described in the foregoing embodiments, the present invention also provides a protein structure diffusion generation and optimization system, comprising: The feature representation building module obtains the amino acid sequence of the protein and transforms it to extract sequence features. The sequence feature representation is used to describe the contextual semantics, structural preferences and evolutionary conservation information of the amino acid sequence, including edge features and node features. The updated vector prediction module combines the sequence feature representation with the spherical harmonic function of the equivariant neural network. Integrate it into a convolutional neural network to construct a convolutional module for predicting the isotropic 3D coordinates of the current protein; The forward prediction update module applies the Adams-Bashforth multi-step method to perform noisy defolding operations on the protein conformation, executes the forward prediction process, and outputs the design latent code. The reverse correction update module performs latent code space operations on the designed latent code, applies the Adams-Moulton multi-step method for reverse denoising and folding operations, performs the reverse generation process, and outputs the predicted conformation. The bidirectional result coupling module uses a bidirectional explicit linear multi-step method to perform bidirectional correction on the forward prediction process and the backward generation process, and constructs a diffusion generation model for protein structures. The structure result generation module, based on the convolution module and the diffusion generation model, obtains the final three-dimensional structure result of the protein.

[0045] The system described above in this invention can effectively realize the protein structure diffusion generation and optimization method, and the technical effects it can achieve are as described in the above embodiments, which will not be repeated here.

[0046] As a preferred embodiment of the above, the bidirectional result coupling module includes: The weighted fusion unit performs weighted fusion of the forward explicit linear multistep prediction update result and the reverse explicit linear multistep correction update result according to the preset bidirectional fusion weights to obtain the coupled structure state at the corresponding time step. The state sequence acquisition unit performs weighted fusion sequentially along the time step sequence to obtain the bidirectional coupled structure state sequence of the entire coverage diffusion generation process; The diffusion model generation unit, based on the bidirectional coupled structural state sequence, determines the model parameters used to describe the evolution of the three-dimensional structural state of proteins, and obtains the diffusion generation model based on the model parameters.

[0047] Similarly, the above-mentioned optimization schemes for the system can also achieve the optimization effects corresponding to the methods in Embodiment 1, which will not be repeated here.

[0048] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for protein structure diffusion generation and optimization, characterized in that, The method includes: The amino acid sequence of the protein is obtained and transformed to extract sequence features, which represent contextual semantics, structural preferences and evolutionary conservation information used to describe the amino acid sequence, including edge features and node features; The sequence feature representation is compared with the spherical harmonic function of the equivariant neural network. Integrate it into a convolutional neural network to construct a convolutional module for predicting the isotropic 3D coordinates of the current protein; The Adams-Bashforth multi-step method is applied to the protein conformation to perform a noisy defolding operation, a forward prediction process is executed, and the design latent code is output. Perform latent code space operations on the designed latent code, apply the Adams-Moulton multi-step method for inverse denoising and folding operations, perform inverse generation process, and output predicted conformation; A bidirectional explicit linear multistep method is used to bidirectionally correct the forward prediction process and the reverse generation process to construct a diffusion generation model for protein structures. Based on the convolution module and the diffusion generation model, the final three-dimensional structure of the protein is obtained.

2. The protein structure diffusion generation and optimization method according to claim 1, characterized in that, The sequence feature representation is compared with the spherical harmonic function of the equivariant neural network. Integrating into a convolutional neural network, a convolutional module is constructed to predict the isotropic 3D coordinates of the current protein, including: The sequence feature representation is subjected to feature mapping processing to obtain a high-dimensional feature representation for convolution operations; Based on the relative spatial relationships between protein residues, the corresponding spatial geometric information is calculated, and a spherical harmonic function representation is generated based on the spatial geometric information. The high-dimensional feature representation is fused with the spherical harmonic function representation to obtain a fused feature representation that satisfies spatial rotation equivariance and translation invariance. Based on the fusion feature representation, perform equivariant convolution operation to extract spatial features related to the current three-dimensional structural state of the protein while maintaining the equivariance of three-dimensional space. Based on the result of the isovariant convolution operation, the isovariant three-dimensional coordinate update amount corresponding to the current three-dimensional structural state of the protein is predicted.

3. The protein structure diffusion generation and optimization method according to claim 1, characterized in that, The convolutional neural network includes: An embedding layer is used to receive the protein structural feature representation and map the protein structural feature representation to a high-dimensional feature space; The geometric feature processing layer is used to extract spatial geometric features and calculate spherical harmonic function representations based on the relative positional relationships between residues in a three-dimensional spatial coordinate system. The equivariant convolutional layer performs joint feature extraction on the protein structure feature representation and the spatial geometric features through equivariant convolution and tensor product operations, while maintaining the equivariance of features under spatial rotation and translation transformations. The output layer is used to predict the structural state update vector corresponding to the three-dimensional structural state of the protein based on the feature representation output by the isovariant convolutional layer.

4. The protein structure diffusion generation and optimization method according to claim 1, characterized in that, After performing the forward explicit linear multistep prediction update, the method also includes attribute interpolation, directional editing, and condition constraint processing on the structural latent code representation.

5. The protein structure diffusion generation and optimization method according to claim 1, characterized in that, The forward explicit linear multistep prediction update includes: Obtain the three-dimensional structural state at multiple historical time steps and its corresponding structural state update vector; The predicted update amount for the current time step is obtained by linearly combining the structure state update vectors of the multiple historical time steps. The predicted update amount is numerically synthesized with the three-dimensional structural state of the most recent time step, and the three-dimensional structural state of the most recent time step is updated to obtain the structural latent code representation corresponding to the current time step.

6. The protein structure diffusion generation and optimization method according to claim 1, characterized in that, The reverse explicit linear multistep correction update includes: The structural latent code is used as the starting state for reverse derivation; On the timeline, in order of time steps earlier than the current time step, obtain the three-dimensional structural state and the corresponding structural state update vector for multiple reverse historical time steps; Based on the three-dimensional structural states of the multiple reverse historical time steps and the corresponding structural state update vectors, the three-dimensional structural state of the protein in the current reverse time step is corrected and updated. The three-dimensional structure state is gradually restored from back to front by sequentially traversing each reverse time step along the time axis and performing the correction update at each reverse time step.

7. The protein structure diffusion generation and optimization method according to claim 1, characterized in that, A diffusion generation model for protein structures is constructed by bidirectionally correcting the forward prediction process and the reverse generation process using a bidirectional explicit linear multistep method, including: According to the preset bidirectional fusion weights, the forward explicit linear multistep prediction update result and the reverse explicit linear multistep correction update result are weighted and fused to obtain the coupled structure state at the corresponding time step. The weighted fusion is performed sequentially along the time step sequence to obtain a bidirectional coupled structural state sequence covering the entire diffusion generation process; Based on the bidirectional coupled structural state sequence, model parameters for describing the evolution of the three-dimensional structural state of the protein are determined, and the diffusion generation model is obtained based on the model parameters.

8. The protein structure diffusion generation and optimization method according to claim 1, characterized in that, When bidirectionally coupling the forward explicit linear multistep prediction update result and the reverse explicit linear multistep correction update result, the method further includes assigning adjustable fusion weights to both, and adaptively adjusting the fusion weights according to the time step position of the three-dimensional structure state.

9. A protein structure diffusion generation and optimization system, characterized in that, The system includes: The feature representation construction module obtains the three-dimensional structural state of the protein and constructs a protein structural feature representation for diffusion generation, wherein the feature representation is used to describe the conformational state of the protein in space. The update vector prediction module inputs the protein structure feature representation into a convolutional neural network to predict the structure state update vector corresponding to the current three-dimensional structure state of the protein. The forward prediction update module, based on the structure state update vector, performs forward explicit linear multi-step prediction update on the three-dimensional structure state based on multiple historical time steps, extrapolates and evolves the three-dimensional structure state, and obtains the corresponding structure latent code representation. The reverse correction update module uses the structural latent code representation as the initial state and reconstructs the three-dimensional structural state in reverse along the time axis through reverse explicit linear multi-step correction update based on the structural state of multiple historical time steps. The bidirectional result coupling module is based on the bidirectional coupling of forward explicit linear multistep prediction results and reverse explicit linear multistep correction update results to construct a diffusion generation model of protein structure. The structure result generation module, based on the diffusion generation model, obtains the final three-dimensional structure result of the protein.

10. The protein structure diffusion generation and optimization system according to claim 9, characterized in that, The bidirectional result coupling module includes: The result weighted fusion unit performs weighted fusion of the forward explicit linear multistep prediction update result and the reverse explicit linear multistep correction update result according to the preset bidirectional fusion weights to obtain the coupled structure state at the corresponding time step. The state sequence acquisition unit performs the weighted fusion sequentially along the time step sequence to obtain the bidirectional coupled structure state sequence of the entire coverage diffusion generation process; The diffusion model generation unit determines model parameters to describe the evolution of the three-dimensional structural state of the protein based on the bidirectional coupled structural state sequence, and obtains the diffusion generation model based on the model parameters.

Images