A secondary structure-guided graph diffusion protein prediction method and system

Abstract

This invention relates to the field of protein structure prediction technology, and particularly to a method and system for predicting diffusion proteins guided by secondary structure planning. The method includes: acquiring residue feature representations of the amino acid sequence of the protein to be predicted; determining the secondary structure type corresponding to each residue to obtain secondary structure prediction results; constructing a set of geometric constraints to characterize the spatial relationships between residues and determining the constraint weights corresponding to each geometric constraint; initializing the protein's three-dimensional structure generation process and introducing constraint guidance information to generate an initial protein three-dimensional structure; segmenting the amino acid sequence with overlapping regions to obtain multiple substructures and fusing them based on the overlapping regions; and post-processing and optimizing the initial protein three-dimensional structure or the fused structure to obtain the three-dimensional structure prediction results. This invention effectively solves the problem of insufficient prediction quality and controllability caused by the lack of effective guidance in structure generation in existing protein structure prediction methods.

Description

Technical Field

[0001] This invention relates to the field of protein structure prediction technology, and in particular to a method and system for predicting secondary structure-guided diffusion proteins. Background Technology

[0002] Protein structure prediction refers to the process of inferring the spatial conformation of a protein based on its amino acid sequence information using computational models. With the continuous accumulation of protein sequence data, obtaining reliable three-dimensional structures from sequences efficiently and accurately has become a key technical challenge in structural biology and computational biology. Existing techniques include: one category of methods is structure prediction based on physical simulation or statistical sampling; another category is based on templates or homology alignment; and yet another category is end-to-end prediction methods based on deep learning, which directly learn the mapping relationship between protein sequences and structures through neural networks. Furthermore, diffusion-based protein structure generation methods have been developed on this basis, generating target structures through progressive denoising.

[0003] Existing end-to-end prediction methods and conventional diffusion generation methods typically focus on data-driven models directly outputting results. While these methods can achieve good prediction performance, their structure generation process often lacks clear intermediate planning and effective guidance, leading to insufficient utilization of prior structural information. This is particularly true in protein structure prediction, where secondary structure information is closely linked to the final 3D conformation. However, in current technologies, this information is often treated merely as input features or auxiliary conditions, failing to be further transformed into a guiding mechanism that substantially constrains the 3D structure generation process. This easily results in insufficient interpretability, inadequate guidance, and insufficient controllability of the results.

[0004] The core flaw of existing technologies lies in the lack of effective guidance in the protein three-dimensional structure generation process, which makes it difficult for prior information such as secondary structure to truly influence the direction of structure generation, thereby affecting the prediction quality and the controllability of generation.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the general background of this disclosure and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] This invention provides a method and system for predicting secondary structure-guided diffusion proteins, which can effectively solve the problems in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting secondary structure-guided diffusion proteins, the method comprising: Obtain the amino acid sequence of the protein to be predicted, and extract residue feature representations based on the amino acid sequence; Based on the residue feature representation, the secondary structure type corresponding to each residue in the amino acid sequence is determined, and the secondary structure prediction result is obtained. Based on the secondary structure prediction results, a set of geometric constraints is constructed to characterize the spatial relationship between residues, and the constraint weights corresponding to each geometric constraint in the set of geometric constraints are determined. The protein three-dimensional structure generation process is initialized based on the amino acid sequence, and constraint guidance information generated by the geometric constraint set is introduced during the back sampling process to generate the initial protein three-dimensional structure. Based on the sampling stage information of the reverse sampling process and the satisfaction of the geometric constraint set by the initial protein three-dimensional structure, the guidance intensity of the constraint guidance information is dynamically adjusted. When the length of the amino acid sequence meets the preset condition, the amino acid sequence is segmented with overlapping regions, and each segment is subjected to a three-dimensional structure generation process guided by the geometric constraint set and with dynamically adjustable guidance intensity, so as to obtain multiple substructures and fuse the multiple substructures based on the overlapping regions. The initial three-dimensional structure of the protein or the structure obtained by fusing multiple substructures is post-processed and optimized to obtain the predicted three-dimensional structure of the protein to be predicted.

[0008] Furthermore, residue feature representation is extracted based on amino acid sequences, including: Obtain the sequence information of each amino acid residue in the protein to be predicted, and encode the sequence information into a corresponding sequence representation; The sequence representation is input into a protein language model to extract context-aware features corresponding to each amino acid residue; Output the embedding features corresponding to each amino acid residue in the amino acid sequence as the residue feature representation.

[0009] Furthermore, the secondary structure prediction results are obtained, including: The residue feature representation is read, and secondary structure classification is performed for each residue separately; Based on the secondary structure classification process, prediction results are obtained for each residue belonging to different secondary structure categories; Based on the prediction results corresponding to each residue, the secondary structure type of each residue in the amino acid sequence is determined, and the secondary structure prediction result corresponding to the amino acid sequence is formed; The predicted secondary structure results are classified into three structural types: α-helix, β-fold, and random coil.

[0010] Further, the geometric constraint set is generated and the weights are determined, including: Based on the secondary structure prediction results, peptide bond distance constraints are generated for adjacent residues; Based on the continuous α-spiral regions in the secondary structure prediction results, a spiral internal distance constraint is generated; Based on the continuous β-fold regions in the secondary structure prediction results, generate fold internal distance constraints; The peptide bond distance constraint, the helical interior distance constraint, and the fold interior distance constraint are uniformly constructed into the geometric constraint set; The weights of each geometric constraint in the geometric constraint set are determined based on the confidence information corresponding to the secondary structure prediction results and the prior reliability of the constraint types.

[0011] Further, the initial three-dimensional structure of the protein is generated, including: The three-dimensional structure of the protein is represented as a set of spatial coordinates corresponding to each residue; Based on the set of geometric constraints, a constraint loss is constructed that reflects the deviation between the current structure and the target geometric relationship; The adjustment direction of the current structural coordinates relative to the set of geometric constraints is calculated based on the constraint loss. During the backsampling process, the adjustment direction is introduced as the constraint guidance information into the structural state update process. Based on the constraint guidance information, the structural state of each sampling step in the reverse sampling process is corrected to generate the initial three-dimensional protein structure.

[0012] Furthermore, dynamically adjusting the guidance strength of the constraint guidance information includes: Obtain the current sampling stage information corresponding to the reverse sampling process; Determine the time-related guidance intensity scheduling value corresponding to the sampling process based on the current sampling phase information; The adaptive error adjustment value is determined based on the current structure's satisfaction of the geometric constraint set. The dynamic guidance strength of the current sampling step is determined based on the time-related guidance strength scheduling value and the error adaptive adjustment value. The dynamic guidance intensity is used to adjust the influence of the constraint guidance information in the subsequent backsampling process.

[0013] Furthermore, multiple substructures are obtained and merged, including: When the length of the amino acid sequence meets a preset condition, the amino acid sequence is divided into multiple subsequences having the overlapping region; The three-dimensional structure generation process is performed on each of the sub-sequences to obtain the sub-structure corresponding to each of the sub-sequences; Identify multiple predicted coordinates of the same residue in different substructures; Based on the positional relationship between the substructures within the overlapping region, the multiple predicted coordinates are weighted and fused. The complete three-dimensional structure of the protein is generated based on the coordinates of each fused residue.

[0014] Furthermore, the predicted three-dimensional structure of the protein was obtained, including: Construct a composite energy function consisting of distance constraints, angle constraints, and spatial conflict terms; The distance constraint term is calculated based on the set of geometric constraints to characterize the degree to which the current structure satisfies the geometric distance constraints; The angle constraint term is calculated based on the deviation between the actual value and the reference value of the dihedral angle of the main chain of each residue. The spatial conflict term is calculated based on the relationship between the actual distance between atoms and the minimum allowable distance to characterize unreasonable close-range contact in the current structure; The initial three-dimensional structure of the protein or the structure formed by the fusion of the substructures is iteratively optimized based on the composite energy function. Output the predicted three-dimensional structure of the protein after the iterative optimization.

[0015] A secondary structure-guided planning diffusion protein prediction system, the system comprising: The residue representation module obtains the amino acid sequence of the protein to be predicted and extracts residue feature representations based on the amino acid sequence; The secondary structure module, based on residue feature representation, determines the secondary structure type corresponding to each residue in the amino acid sequence and obtains the secondary structure prediction results; The geometric constraint module constructs a set of geometric constraints to characterize the spatial relationships between residues based on the secondary structure prediction results, and determines the constraint weights corresponding to each geometric constraint in the set of geometric constraints. The initial generation module initializes the protein three-dimensional structure generation process based on the amino acid sequence and introduces constraint guidance information generated by the geometric constraint set during the back sampling process to generate the initial protein three-dimensional structure. The constraint adjustment module dynamically adjusts the guidance intensity of the constraint guidance information based on the sampling stage information of the backsampling process and the satisfaction of the geometric constraint set by the initial three-dimensional protein structure. The fusion generation module, when the length of the amino acid sequence meets the preset conditions, segments the amino acid sequence with overlapping regions, and performs three-dimensional structure generation processing on each segment guided by a geometric constraint set with dynamically adjustable guidance intensity, to obtain multiple substructures and fuse the multiple substructures based on the overlapping regions. The 3D prediction module performs post-processing optimization on the initial 3D structure of the protein or the structure obtained by fusing multiple substructures to obtain the 3D structure prediction result of the protein to be predicted.

[0016] Further, the residue representation module includes: The sequence coding unit obtains the sequence information of each amino acid residue in the protein to be predicted and encodes the sequence information into the corresponding sequence representation; The context-aware unit inputs the sequence representation into the protein language model to extract the context-aware features corresponding to each amino acid residue; The feature embedding unit outputs the embedding features corresponding to each amino acid residue in the amino acid sequence, which serve as residue feature representations.

[0017] The technical solution of this invention can achieve the following technical effects: By first predicting secondary structures based on protein sequences and generating corresponding geometric constraints, and then using these geometric constraints to dynamically guide the three-dimensional structure diffusion generation process, and combining long sequence segmentation parallel generation and post-processing optimization to obtain the final protein structure, the problem of insufficient prediction quality and controllability caused by the lack of effective guidance in structure generation in existing protein structure prediction is effectively solved.

[0018] 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

[0019] 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.

[0020] Figure 1 This is a flowchart illustrating a secondary structure-guided planning method for predicting protein diffusion. Figure 2 This is a flowchart illustrating the process of secondary structure prediction. Figure 3A flowchart illustrating the process of establishing geometric constraints; Figure 4 A schematic diagram of the planning diffusion sampling process; Figure 5 A flowchart for post-processing optimization. Detailed Implementation

[0021] 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.

[0022] 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.

[0023] Example 1; like Figure 1 As shown, this application provides a method for predicting secondary structure-guided planning diffusion proteins, the method comprising: S10: Obtain the amino acid sequence of the protein to be predicted, and extract residue feature representations based on the amino acid sequence; S20: Based on residue feature representation, determine the secondary structure type corresponding to each residue in the amino acid sequence and obtain the secondary structure prediction results; S30: Based on the secondary structure prediction results, construct a set of geometric constraints to characterize the spatial relationship between residues, and determine the constraint weights corresponding to each geometric constraint in the set of geometric constraints; S40: The protein three-dimensional structure generation process is initialized based on the amino acid sequence, and constraint guidance information generated by the geometric constraint set is introduced during the back sampling process to generate the initial protein three-dimensional structure; S50: Based on the sampling stage information of the reverse sampling process and the satisfaction of the geometric constraint set by the initial protein three-dimensional structure, the guidance intensity of the constraint guidance information is dynamically adjusted. S60: When the length of the amino acid sequence meets the preset conditions, the amino acid sequence is segmented with overlapping regions, and each segment is subjected to three-dimensional structure generation processing guided by a geometric constraint set with dynamically adjustable guidance intensity to obtain multiple substructures and fuse the multiple substructures based on the overlapping regions. S70: Post-processing optimization is performed on the initial three-dimensional structure of the protein or the structure obtained by fusing multiple substructures to obtain the predicted three-dimensional structure of the protein to be predicted.

[0024] Specifically, in a preferred embodiment of the present invention, the amino acid sequence of the protein to be predicted is first obtained, and the amino acid sequence is then standardized and preprocessed. This preprocessing may include validating residue characters, removing or replacing non-standard characters, and generating a residue index according to the sequence order to ensure data consistency in subsequent feature extraction and structure generation processes. Based on this, the amino acid sequence is input into a protein language model to extract context-aware feature representations corresponding to each residue. Preferably, a protein language model capable of representing long-range dependencies is used to encode the entire sequence, so that each residue obtains an embedding feature that integrates upstream and downstream sequence environment information, thereby... This provides structurally biased characterization information for subsequent secondary structure discrimination. Subsequently, secondary structures are classified for each residue based on residue feature representation, preferably into three categories: α-helix, β-sheet, and random coil. Corresponding category confidence information is output simultaneously. In practice, residue segments continuously identified as α-helices can be considered candidate helical regions, residue segments continuously identified as β-sheets can be considered candidate folding regions, and the remaining segments can be considered highly flexible random coil regions, so as to generate geometric constraints consistent with the native folding pattern of proteins. After obtaining the secondary structure prediction results, a system for constructing a representation based on the secondary structure prediction results is further developed. A set of geometric constraints characterizing the spatial relationships between residues is used. This set preferably includes main chain connectivity constraints between adjacent residues, characteristic distance constraints within continuous α-helical regions, and extended conformation constraints within continuous β-sheet regions. Weights are assigned to each geometric constraint based on the confidence level of the secondary structure prediction results and the reliability differences of different constraint types. Connectivity constraints between adjacent residues are preferably assigned higher weights to ensure main chain continuity; constraints within α-helical regions are preferably assigned the next highest weights to highlight the stable characteristics of regular secondary structures; and constraints within β-sheet regions are assigned weights that reflect their extended geometric properties. This ensures the proper functioning of the secondary structure. The prediction results are no longer just static auxiliary information, but are transformed into active guiding conditions that can directly participate in the subsequent three-dimensional structure generation. In the three-dimensional structure generation stage, a diffusion-based back sampling strategy is preferred for protein conformation generation. First, the initial structural state is established based on the amino acid sequence. Then, during the gradual denoising back sampling process, constraint guidance information generated by the geometric constraint set is introduced. Specifically, the deviation of the current structure from the geometric constraint set can be calculated at each sampling step, and the adjustment direction corresponding to the deviation is superimposed on the current structure update process so that the structural state gradually converges towards the direction of satisfying the geometric constraint set while maintaining the diversity of diffusion generation.To avoid excessive guidance leading to insufficient exploration of the conformation space, or insufficient guidance leading to the ineffective functioning of geometric constraints, this invention further optimizes the dynamic adjustment of the guidance intensity of the constraint guidance information based on the sampling stage information of the backsampling process and the current generated structure's satisfaction with the set of geometric constraints. Specifically, the following optimized strategies can be adopted: maintaining weak guidance in the early stages of sampling to preserve sufficient exploration of the conformation space; increasing the guidance intensity in the middle stages of sampling to rapidly converge the structure towards the direction conforming to the prior knowledge of the second-order structure; appropriately weakening and refining the guidance in the later stages of sampling to facilitate local geometric fine-tuning and overall convergence stability; simultaneously, adaptive correction can be performed based on the constraint satisfaction error of the current structure. When the constraint satisfaction error of the current structure is detected, the guidance intensity can be adjusted accordingly. When the deviation from the target is large, the guiding effect is enhanced accordingly; conversely, when the current structure is detected to have adequately met the constraints, the guiding effect is reduced accordingly, thus achieving a dynamic balance between exploration capability and constraint satisfaction. This is one of the important improvements of this invention compared to existing methods that only use fixed-condition injection. For longer protein sequences, it is preferable to segment the amino acid sequence with overlapping regions when the amino acid sequence length reaches a preset condition. Segmentation preferably avoids the middle of continuous regular secondary structures as much as possible to reduce damage to the integrity of substructures. Each segment undergoes a three-dimensional structure generation process guided by a set of geometric constraints, with the guiding intensity dynamically adjustable, thereby obtaining multiple substructures. Then, the overlapping regions are used to fuse the multiple substructures, preferably using... A weighted approach that smoothly varies along overlapping regions fuses the predicted coordinates of the same residue in different substructures. This allows residues at the center of the overlapping region to inherit more stable predictions of the local substructure, while achieving a natural transition at adjacent boundaries. This avoids structural breaks and geometric distortions caused by simple splicing. For example, in a long protein sequence, if there is an overlapping region between the beginning and middle segments, the predicted coordinates of the same residues within the overlapping region in the beginning and middle substructures can be smoothly weighted to obtain a continuous and naturally transitioning overall structure. After obtaining the initial three-dimensional protein structure or the overall structure obtained by fusing multiple substructures, further post-processing optimization is performed to preferentially construct a structure that simultaneously incorporates distance constraints, angle constraints, and... The composite optimization objective of spatial conflict constraints is to refine the structural coordinates using an iterative optimization method. Distance constraints are used to maintain the target spatial relationships of key residue pairs in the geometric constraint set; angle constraints are used to make the main chain conformation more consistent with the stereochemical rules corresponding to different secondary structures; and spatial conflict constraints are used to eliminate unreasonable close-range contacts between atoms, thereby improving the geometric rationality, physical feasibility, and biochemical rationality of the generated structure. Preferably, an adaptive gradient optimization strategy can be used to gradually update the structural coordinates, and during the iteration process, the system continuously evaluates whether there are problems such as abnormal bond lengths, unreasonable main chain twisting, or atomic collisions in the local structure, until the preset convergence condition is met, at which point the final predicted three-dimensional protein structure is output.Through the above preferred embodiments, this invention achieves a process starting from the amino acid sequence, first forming secondary structure predictions, then converting the secondary structure predictions into geometric constraints, and implementing dynamic guidance during the three-dimensional structure diffusion generation process. Combined with long sequence segmentation and parallel fusion processing, as well as post-processing optimization, the invention ultimately obtains protein three-dimensional structure prediction results that are interpretable, controllable, and structurally reasonable.

[0025] The technical solution of this invention first predicts the secondary structure based on the protein sequence and generates corresponding geometric constraints. Then, the geometric constraints are used to dynamically guide the three-dimensional structure diffusion generation process. Combined with long sequence segmentation parallel generation and post-processing optimization, the final protein structure is obtained. This effectively solves the problem of insufficient prediction quality and controllability caused by the lack of effective guidance in structure generation in existing protein structure prediction.

[0026] Furthermore, residue feature representation based on amino acid sequence extraction includes: Obtain the sequence information of each amino acid residue in the protein to be predicted, and encode the sequence information into the corresponding sequence representation; The sequence representation is input into a protein language model to extract context-aware features corresponding to each amino acid residue; The embedding features corresponding to each amino acid residue in the amino acid sequence are output as residue feature representations.

[0027] As a preferred embodiment of the above embodiments, in one embodiment, such as Figure 2 As shown, the method based on The language model of the architecture By combining deep learning technology, we can achieve [the following]: -spiral, - High-precision classification and prediction of three secondary structures: folded, random curled, and non-folded; Input representation: Let the sequence be... ,in Indicates the first The amino acid sequence at each position belongs to a set of 20 standard amino acids; Model prediction process: Protein sequence is predicted through... The language model processes the data to generate a context-aware embedding representation for each residue: in This represents the embedding matrix of the entire sequence. Indicates the first Embedding vector of residues, yes The embedding dimension of the model The model is based on The architecture is able to capture long-range dependencies in a sequence.

[0028] Furthermore, obtaining secondary structure prediction results includes: Read the residue feature representation and perform secondary structure classification processing for each residue separately; Based on the secondary structure classification, the prediction results of each residue belonging to different secondary structure categories were obtained; Based on the prediction results corresponding to each residue, the secondary structure type of each residue in the amino acid sequence is determined, and the secondary structure prediction results corresponding to the amino acid sequence are formed. The secondary structure prediction results are classified into three structural types: α-helix, β-fold, and random coil.

[0029] As a preferred embodiment of the above embodiments, in one embodiment, for each residue Based on its embedding vector Predicting secondary structure type: in, It is a learnable weight matrix; It is a bias vector; output Indicates the predicted secondary structure type, respectively representing -spiral, - Folding and random curling; The module outputs a complete sequence of secondary structures: Each of them It is obtained by taking the category with the highest probability: .

[0030] Furthermore, generating the set of geometric constraints and determining the weights includes: Based on the secondary structure prediction results, peptide bond distance constraints are generated for adjacent residues; Based on the continuous α-spiral regions in the secondary structure prediction results, a spiral internal distance constraint is generated; The internal distance constraint of the fold is generated based on the continuous β-fold region in the secondary structure prediction results; The peptide bond distance constraint, the helical interior distance constraint, and the fold interior distance constraint are unified into a set of geometric constraints. Based on the confidence information corresponding to the secondary structure prediction results and the prior reliability of the constraint type, the weight of each geometric constraint in the geometric constraint set is determined.

[0031] As a preferred embodiment of the above embodiments, in one embodiment, such as Figure 3As shown, the secondary structure prediction information is transformed into explicit three-dimensional geometric constraints. Based on the protein secondary structure prediction results, mathematical constraints describing the spatial relationships between residues are generated, providing key geometric guidance information for subsequent three-dimensional structure generation. Peptide bond distance constraints: Describe the relationship between adjacent amino acid residues The standard distance constraint of atoms reflects the basic geometric characteristics of the protein backbone and is equal to the distance between all adjacent residue pairs. ,in : peptide bond Typical interatomic distance is is one of the most stable and conserved geometric parameters in the protein backbone, and its weight is usually set to the highest to reflect its high reliability. - Helical internal distance constraint: In the prediction - A continuous region of the spiral, describing the interior of the spiral. inter-residue Hydrogen bond distance constraints between atoms. Formation conditions: For all residue pairs that satisfy the above conditions : In the helix, the carbonyl oxygen of each residue forms a hydrogen bond with the amide hydrogen of the fourth residue downstream; correspondingly The distance is approximately The spiral ascent distance is approximately / residue; - Folding internal distance constraint: In a continuous region predicted as β-folds, describe the interior of the folds. inter-residue Atom stretching distance constraint E, generation conditions: For all residue pairs that satisfy the above conditions : - The folds exhibit an extended conformation, with longer distances between residues; reflecting... - The extension geometry of the folded chain provides spatial constraints for the formation of the folded structure; Constraint set definition and weight calculation: All generated geometric constraints are uniformly represented as constraint geometry. : in Represents residues The target distance between them; This represents the weighting coefficient of the constraint, reflecting the confidence level of the constraint; Assign values ​​based on the confidence level of the secondary structure prediction: in, It is a residue The probability of being predicted as the corresponding secondary structure type. This is the constraint type factor, which reflects the prior reliability of different types of constraints. The factor is 1.0 for peptide bond constraints, 0.8 for helical constraints, and 0.6 for folding constraints.

[0032] Furthermore, generating the initial three-dimensional structure of the protein includes: The three-dimensional structure of a protein is represented as a set of spatial coordinates corresponding to each residue. Construct a constraint loss based on a set of geometric constraints to reflect the deviation between the current structure and the target geometric relationship; Calculate the adjustment direction of the current structural coordinates relative to the set of geometric constraints based on the constraint loss; During the backsampling process, the adjustment direction is used as constraint guidance information and introduced into the structural state update process; Based on constraint-guided information, the structural state of each sampling step in the backsampling process is corrected to generate the initial three-dimensional protein structure.

[0033] As a preferred embodiment of the above embodiments, in one embodiment, such as Figure 4 As shown, three-dimensional protein structures are generated stepwise from random noise. An improved diffusion model framework is used to introduce a secondary structure guidance mechanism in the standard denoising process. Combined with dynamic intensity control and parallel sampling strategies, efficient and accurate protein structure generation is achieved. Standard diffusion model basics: The theoretical foundation mentions the standard diffusion model, and the denoising update formula is: in Indicates from the first The first diffusion step backpropagation yields the... The basic structural state of each diffusion step; Indicates the current number Protein structural states under each diffusion step; Indicates the first The signal retention ratio corresponding to each diffusion step ,in Indicates the first The noise addition ratio for each diffusion step; Indicates that the parameter is The denoising network takes the current state as input. and current number of steps The predicted noise value at the output; Represents the model parameters of the denoising network; Indicates the first The variance term or random perturbation strength of each diffusion step; This represents a random noise variable sampled from a standard distribution; Indicates from step 1 to step 2. Step by step, all The cumulative product, ,in This represents the signal retention ratio corresponding to the s-th diffusion step in the cumulative product, where s represents the traversal step index used when calculating the cumulative product; Planning diffusion sampling: That is, a secondary structure guiding term is introduced into the standard diffusion process, and a geometric constraint loss function is defined: in, and Represents the coordinates of the i-th and j-th residues in the current protein structure; Calculate the constraint gradient: in, This represents the structural state at the current t-th diffusion step. Find the gradient; The updated formula for planning diffusion denoising is: in It is a dynamic guiding intensity function. This represents the guiding strength at the t-th sampling step; Parallel sampling strategy: For long amino acid sequences ( ), split the sequence into Overlapping subsequences: The length of the overlapping region is ; The parallel sampling strategy involves independently planning and spreading sampling for each subsequence to obtain the substructure. ; Block merging uses a weighted average method to merge substructures. For the first substructure in the sequence... The coordinates of a residue are the weighted sum of the predicted coordinates of all blocks that include that residue: in It is a set of blocks containing the i-th residue; weight The weight changes linearly along the sequence position within the overlapping region of block k. When it is located at the center of the overlapping region of the block, the weight is 1, and when it is located at the boundary with the adjacent block, the weight drops to 0, thereby achieving a smooth transition.

[0034] Furthermore, dynamically adjusting the guidance strength of constraint guidance information includes: Obtain the current sampling stage information corresponding to the backsampling process; Determine the time-related guidance intensity scheduling value corresponding to the sampling process based on the information of the current sampling phase. The adaptive error adjustment value is determined based on the current structure's satisfaction of the set of geometric constraints. The dynamic guidance strength of the current sampling step is determined by combining the time-related guidance strength scheduling value and the error adaptive adjustment value. The influence of constraint guidance information in the subsequent backsampling process is adjusted by using dynamic guidance intensity.

[0035] As a preferred embodiment of the above embodiments, in one embodiment, firstly, the intensity scheduling of the time-dependent diffusion model step size is defined, and a progress variable is defined: in, This represents the total number of steps in the entire diffusion sampling process; Define the dynamic intensity function as follows: in, Represents the maximum guiding strength of the dynamic intensity function, scheduling function Piecewise linear design is adopted: Then, the intensity of the error adaptive adjustment is achieved by introducing an error adaptive factor: in This is a preset reference loss value used for normalization; the final dynamic strength is: in This indicates a time-scheduling item.

[0036] Furthermore, multiple substructures are obtained and merged, including: When the length of the amino acid sequence meets the preset conditions, the amino acid sequence is divided into multiple subsequences with overlapping regions; Perform three-dimensional structure generation processing on each subsequence to obtain the substructure corresponding to each subsequence; Identify multiple predicted coordinates of the same residue in different substructures; Based on the positional relationship of each substructure within the overlapping area, multiple predicted coordinates are weighted and fused. The complete three-dimensional structure of the protein is generated based on the coordinates of each residue after fusion.

[0037] As a preferred embodiment of the above embodiments, in a preferred embodiment of the present invention, when the amino acid sequence length of the protein to be predicted reaches a preset condition, the existing method of mechanically segmenting and simply splicing long sequences is no longer used. Instead, the amino acid sequence is preferably segmented with overlapping regions based on the secondary structure prediction results obtained in the preceding steps, so that the segmentation position avoids the interior of continuous regular secondary structures as much as possible, especially avoiding truncating continuous α-helix segments or β-sheet segments from the middle, so as to reduce the damage to the integrity of local folding units by the segmentation boundary; specifically, the continuous distribution of secondary structure categories can be scanned along the sequence first, in While meeting the single-segment length requirement, segment boundaries are preferentially set near random curled segments or secondary structure transition regions, and a certain length of overlap is retained between adjacent subsequences to ensure that residues on both sides of the boundary are preserved in both subsequences, facilitating a smooth transition during subsequent fusion. After segmentation, three-dimensional structure generation processing is performed on each subsequence independently. This involves implementing geometrically constrained, dynamically adjustable planning and diffusion sampling for each subsequence to obtain its corresponding substructure. In practice, each subsequence can be executed in parallel on the same computing node using multi-threading or multi-processing, or mapped to different... The process can be executed independently on graphics processing devices or different computing units, as long as it can ensure that the structure generation of each subsequence is completed under the same geometric constraint guidance principle. Preferably, for each subsequence, the geometric constraint set constructed from the secondary structure prediction results in the previous step is still used, and the constraint guidance direction and dynamic guidance intensity are calculated according to the range of residues contained in the subsequence during their respective backsampling process, so that each substructure has both local integrity and consistency with the global prior. After obtaining multiple substructures, the multiple predicted coordinates of the same residue in different substructures are first identified, preferably by the one-to-one correspondence of residue numbers. That is, for residues located in the overlapping region, their predicted positions in the preceding and following substructures are extracted simultaneously, and a coordinate correspondence relationship of residues with the same name is established; then, the multiple predicted coordinates are weighted and fused according to the positional relationship of each substructure in the overlapping region. Preferably, a linear weight allocation method that gradually changes along the position of the overlapping region is adopted, so that the coordinates of residues located near the center of the overlapping region of a certain substructure retain more of the prediction results of that substructure, while the coordinates of residues located near the adjacent boundary gradually reduce the weight of that substructure and correspondingly increase the weight of the adjacent substructure, thereby realizing a natural transition from one substructure to another.

[0038] Furthermore, the predicted three-dimensional structure of the protein is obtained, including: Construct a composite energy function consisting of distance constraints, angle constraints, and spatial conflict terms; Distance constraint terms are calculated based on the set of geometric constraints to characterize the degree to which the current structure satisfies the geometric distance constraints; The angle constraint term is calculated based on the deviation between the actual value and the reference value of the dihedral angle of each residue's main chain. Spatial conflict terms are calculated based on the relationship between the actual distance between atoms and the minimum allowable distance to characterize unreasonable close-range contact in the current structure; Iterative optimization of the initial three-dimensional structure of the protein or the structure formed by the fusion of substructures is performed based on the composite energy function; Output the predicted three-dimensional structure of the protein after iterative optimization.

[0039] As a preferred embodiment of the above embodiments, in one embodiment, such as Figure 5 As shown, the preliminary structure generated by the planned diffusion sampling is finely adjusted. By optimizing the composite energy function, which includes distance constraints, angle constraints, and spatial conflict constraints, the geometric rationality, physical rationality, and biochemical feasibility of the protein structure are further improved, providing quality assurance for the final structure output. Composite energy function: The goal of post-processing optimization is to minimize the composite energy function. The function consists of three main parts: in, These are distance constraint terms that ensure the structure meets geometric distance constraints. It is an angle constraint term that optimizes the dihedral angle distribution of the main chain; Spatial conflict terms eliminate unreasonable collisions between atoms; , These are weighting coefficients, used to balance the importance of various constraints; Distance constraint terms measure the current structure With geometric constraints Set conformity: in, , These are residues and of coordinate; It is the target distance (e.g., peptide bond 3.8 Å, helix). wait); These are constraint weights, reflecting the confidence level of the constraints; Angle constraints: Angle constraint terms optimize the dihedral angle distribution of the main chain, making the structure more consistent with the stereochemical rules of proteins: in , These are residues The actual value of the dihedral angle of the main chain is determined by four adjacent... Atomic coordinates were used for calculation; , It is based on the dihedral angle and the type setting of the secondary structure. - Spiral region , , - Folded area , The reference value for the irregular curled region is based on the allowable region of the Laplace chart; Spatial conflict items: Spatial conflict terms eliminate unreasonable close-range contact between atoms, ensuring the physical feasibility of the structure. in, It is the minimum allowed distance, usually set to When the distance between two atoms is less than A penalty is incurred; the penalty is proportional to the square of the insufficient distance. Adam optimizer: Iterative optimization is performed using the Adam optimizer. Adam combines the advantages of the momentum method and RMSProp, featuring adaptive learning rate and momentum term, making it suitable for non-convex optimization problems. The iterative update formula is as follows: in, It is the first The structural coordinates of the next iteration; It is the learning rate, which controls the update step size; Is the energy function in The gradient at that point.

[0040] Example 2; Based on the same inventive concept as the secondary structure-guided planning diffusion protein prediction method described in the foregoing embodiments, the present invention also provides a secondary structure-guided planning diffusion protein prediction system, the system comprising: The residue representation module obtains the amino acid sequence of the protein to be predicted and extracts residue feature representations based on the amino acid sequence; The secondary structure module, based on residue feature representation, determines the secondary structure type corresponding to each residue in the amino acid sequence and obtains the secondary structure prediction results; The geometric constraint module constructs a set of geometric constraints to characterize the spatial relationships between residues based on the secondary structure prediction results, and determines the constraint weights corresponding to each geometric constraint in the set of geometric constraints. The initial generation module initializes the protein three-dimensional structure generation process based on the amino acid sequence and introduces constraint guidance information generated by the geometric constraint set during the back sampling process to generate the initial protein three-dimensional structure. The constraint adjustment module dynamically adjusts the guidance intensity of the constraint guidance information based on the sampling stage information of the backsampling process and the satisfaction of the geometric constraint set by the initial three-dimensional protein structure. The fusion generation module, when the length of the amino acid sequence meets the preset conditions, segments the amino acid sequence with overlapping regions, and performs three-dimensional structure generation processing on each segment guided by a geometric constraint set with dynamically adjustable guidance intensity, to obtain multiple substructures and fuse the multiple substructures based on the overlapping regions. The 3D prediction module performs post-processing optimization on the initial 3D structure of the protein or the structure obtained by fusing multiple substructures to obtain the 3D structure prediction result of the protein to be predicted.

[0041] The system described above in this invention can effectively implement a secondary structure-guided planning diffusion protein prediction method, and the technical effects it can achieve are as described in the above embodiments, and will not be repeated here.

[0042] Furthermore, the residues represent modules, including: The sequence coding unit obtains the sequence information of each amino acid residue in the protein to be predicted and encodes the sequence information into the corresponding sequence representation; The context-aware unit inputs the sequence representation into the protein language model to extract the context-aware features corresponding to each amino acid residue; The feature embedding unit outputs the embedding features corresponding to each amino acid residue in the amino acid sequence, which serve as residue feature representations.

[0043] 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.

[0044] 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 predicting secondary structure-guided diffusion proteins, characterized in that, The method includes: Obtain the amino acid sequence of the protein to be predicted, and extract residue feature representations based on the amino acid sequence; Based on the residue feature representation, the secondary structure type corresponding to each residue in the amino acid sequence is determined, and the secondary structure prediction result is obtained. Based on the secondary structure prediction results, a set of geometric constraints is constructed to characterize the spatial relationship between residues, and the constraint weights corresponding to each geometric constraint in the set of geometric constraints are determined. The protein three-dimensional structure generation process is initialized based on the amino acid sequence, and constraint guidance information generated by the geometric constraint set is introduced during the back sampling process to generate the initial protein three-dimensional structure. Based on the sampling stage information of the reverse sampling process and the satisfaction of the geometric constraint set by the initial protein three-dimensional structure, the guidance intensity of the constraint guidance information is dynamically adjusted. When the length of the amino acid sequence meets the preset condition, the amino acid sequence is segmented with overlapping regions, and each segment is subjected to a three-dimensional structure generation process guided by the geometric constraint set and with dynamically adjustable guidance intensity, so as to obtain multiple substructures and fuse the multiple substructures based on the overlapping regions. The initial three-dimensional structure of the protein or the structure obtained by fusing multiple substructures is post-processed and optimized to obtain the predicted three-dimensional structure of the protein to be predicted.

2. The method for predicting secondary structure-guided diffusion proteins according to claim 1, characterized in that, Residue feature representation based on amino acid sequence extraction includes: Obtain the sequence information of each amino acid residue in the protein to be predicted, and encode the sequence information into a corresponding sequence representation; The sequence representation is input into a protein language model to extract context-aware features corresponding to each amino acid residue; Output the embedding features corresponding to each amino acid residue in the amino acid sequence as the residue feature representation.

3. The method for predicting secondary structure-guided diffusion proteins according to claim 1, characterized in that, Obtain the secondary structure prediction results, including: The residue feature representation is read, and secondary structure classification is performed for each residue separately; Based on the secondary structure classification process, prediction results are obtained for each residue belonging to different secondary structure categories; Based on the prediction results corresponding to each residue, the secondary structure type of each residue in the amino acid sequence is determined, and the secondary structure prediction result corresponding to the amino acid sequence is formed; The predicted secondary structure results are classified into three structural types: α-helix, β-fold, and random coil.

4. The method for predicting secondary structure-guided diffusion proteins according to claim 3, characterized in that, Generate a set of geometric constraints and determine their weights, including: Based on the secondary structure prediction results, peptide bond distance constraints are generated for adjacent residues; Based on the continuous α-spiral regions in the secondary structure prediction results, a spiral internal distance constraint is generated; Based on the continuous β-fold regions in the secondary structure prediction results, generate fold internal distance constraints; The peptide bond distance constraint, the helical interior distance constraint, and the fold interior distance constraint are uniformly constructed into the geometric constraint set; The weights of each geometric constraint in the geometric constraint set are determined based on the confidence information corresponding to the secondary structure prediction results and the prior reliability of the constraint types.

5. The method for predicting secondary structure-guided diffusion proteins according to claim 1, characterized in that, Generate the initial three-dimensional structure of the protein, including: The three-dimensional structure of the protein is represented as a set of spatial coordinates corresponding to each residue; Based on the set of geometric constraints, a constraint loss is constructed that reflects the deviation between the current structure and the target geometric relationship; The adjustment direction of the current structural coordinates relative to the set of geometric constraints is calculated based on the constraint loss. During the backsampling process, the adjustment direction is introduced as the constraint guidance information into the structural state update process. Based on the constraint guidance information, the structural state of each sampling step in the reverse sampling process is corrected to generate the initial three-dimensional protein structure.

6. The method for predicting secondary structure-guided diffusion proteins according to claim 5, characterized in that, Dynamically adjusting the guidance strength of the constraint guidance information includes: Obtain the current sampling stage information corresponding to the reverse sampling process; Determine the time-related guidance intensity scheduling value corresponding to the sampling process based on the current sampling phase information; The adaptive error adjustment value is determined based on the current structure's satisfaction of the geometric constraint set. The dynamic guidance strength of the current sampling step is determined based on the time-related guidance strength scheduling value and the error adaptive adjustment value. The dynamic guidance intensity is used to adjust the influence of the constraint guidance information in the subsequent backsampling process.

7. The method for predicting secondary structure-guided diffusion proteins according to claim 1, characterized in that, Obtain and merge multiple substructures, including: When the length of the amino acid sequence meets a preset condition, the amino acid sequence is divided into multiple subsequences having the overlapping region; The three-dimensional structure generation process is performed on each of the sub-sequences to obtain the sub-structure corresponding to each of the sub-sequences; Identify multiple predicted coordinates of the same residue in different substructures; Based on the positional relationship between the substructures within the overlapping region, the multiple predicted coordinates are weighted and fused. The complete three-dimensional structure of the protein is generated based on the coordinates of each fused residue.

8. The method for predicting secondary structure-guided diffusion proteins according to claim 1, characterized in that, The predicted three-dimensional structure of the protein was obtained, including: Construct a composite energy function consisting of distance constraints, angle constraints, and spatial conflict terms; The distance constraint term is calculated based on the set of geometric constraints to characterize the degree to which the current structure satisfies the geometric distance constraints; The angle constraint term is calculated based on the deviation between the actual value and the reference value of the dihedral angle of the main chain of each residue. The spatial conflict term is calculated based on the relationship between the actual distance between atoms and the minimum allowable distance to characterize unreasonable close-range contact in the current structure; The initial three-dimensional structure of the protein or the structure formed by the fusion of the substructures is iteratively optimized based on the composite energy function. Output the predicted three-dimensional structure of the protein after the iterative optimization.

9. A secondary structure-guided planning diffusion protein prediction system, characterized in that, The system includes: The residue representation module obtains the amino acid sequence of the protein to be predicted and extracts residue feature representations based on the amino acid sequence; The secondary structure module, based on residue feature representation, determines the secondary structure type corresponding to each residue in the amino acid sequence and obtains the secondary structure prediction results; The geometric constraint module constructs a set of geometric constraints to characterize the spatial relationships between residues based on the secondary structure prediction results, and determines the constraint weights corresponding to each geometric constraint in the set of geometric constraints. The initial generation module initializes the protein three-dimensional structure generation process based on the amino acid sequence and introduces constraint guidance information generated by the geometric constraint set during the back sampling process to generate the initial protein three-dimensional structure. The constraint adjustment module dynamically adjusts the guidance intensity of the constraint guidance information based on the sampling stage information of the backsampling process and the satisfaction of the geometric constraint set by the initial three-dimensional protein structure. The fusion generation module, when the length of the amino acid sequence meets the preset conditions, segments the amino acid sequence with overlapping regions, and performs three-dimensional structure generation processing on each segment guided by a geometric constraint set with dynamically adjustable guidance intensity, to obtain multiple substructures and fuse the multiple substructures based on the overlapping regions. The 3D prediction module performs post-processing optimization on the initial 3D structure of the protein or the structure obtained by fusing multiple substructures to obtain the 3D structure prediction result of the protein to be predicted.

10. The secondary structure-guided planning diffusion protein prediction system according to claim 9, characterized in that, The residue representation module includes: The sequence coding unit obtains the sequence information of each amino acid residue in the protein to be predicted and encodes the sequence information into the corresponding sequence representation; The context-aware unit inputs the sequence representation into the protein language model to extract the context-aware features corresponding to each amino acid residue; The feature embedding unit outputs the embedding features corresponding to each amino acid residue in the amino acid sequence, which serve as residue feature representations.

Images