Antibody variable region structure prediction method based on structure prior evaluation

Abstract

The invention relates to the field of artificial intelligence, and particularly discloses an antibody variable region structure prediction method based on structure prior evaluation, which comprises the following steps: S1, screening a high-resolution and high-integrity compound containing a complete antigen and an antibody H / L chain from PDB, and extracting atomic coordinates after processing; s2, calculating geometrical characteristics of a molecular structure and relative spherical polar coordinates of adjacent residues; s3, calculating a structure quality evaluation score according to physical prior and a dihedral angle-atomic spacing mapping formula; s4, screening training data according to scores, and removing low-quality data; s5, constructing an antigen-antibody attribute isomeric graph and initializing a compound graph; s6, embedding related data as enhanced features when the model is initialized; s7, loading an RA-EGN encoder to output a CDR sequence and a 3D coordinate, and splicing a complete antibody structure; and S8, introducing a structure prior loss function fine tuning model, and performing multi-index screening on an optimal result. According to the invention, the problems of low prediction precision and insufficient structure prior utilization of the existing antibody CDR region are solved, and the prediction accuracy of the antibody variable region is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for predicting the structure of antibody variable regions based on prior structural assessment. Background Technology

[0002] Antibodies are Y-shaped proteins produced by the immune system, whose main function is to specifically recognize, bind to, and neutralize antigens such as pathogens. The specificity of antibody-antigen binding is mainly determined by six complementarity-determining regions (CDRs, including HCDR1-3 and LCDR1-3) on its heavy and light chains. Based on this mechanism, designing CDR structures that can precisely target antigenic epitopes has become a key step in the development of therapeutic antibodies, and has important application value in multiple therapeutic areas such as cancer, infectious diseases, and autoimmune diseases.

[0003] However, the theoretical candidate space for CDR sequences exceeds 20. 60 The antigen-antigen pair is highly variable in structure and conformation. Traditional experimental methods, such as phage display and monoclonal antibody preparation, typically require numerous screening and validation steps to obtain candidate molecules, resulting in lengthy processes, high costs, and limited success rates. Particularly during antibody affinity maturation, experimental methods struggle to efficiently and systematically optimize the antibody-antigen binding properties; therefore, computational methods are urgently needed to overcome these technical bottlenecks.

[0004] The development of antibody computation methods has evolved from traditional energy models to deep learning models. Early energy-based models, represented by Adolf-Bryfogle et al. and Lapidoth et al., mainly relied on complex biophysical energy functions to sample and analyze protein sequences and structures. These models had low computational efficiency and were prone to getting trapped in local optima.

[0005] With the continuous improvement of computing power and the advancement of bioinformatics technology, molecular dynamics simulation, computer-aided antibody engineering, and deep learning methods have gradually become mainstream technical approaches. Molecular dynamics simulation can predict binding stability by modeling antibody-antigen interactions, but it is highly dependent on high-performance computing resources and is difficult to apply to the design optimization of unknown antigens, typically used only as an auxiliary screening tool. Computer-aided antibody engineering, by combining algorithms with database information and incorporating experimental data, has improved design efficiency to some extent; however, its overall performance still highly depends on the quality of the training data and the level of algorithm optimization.

[0006] In recent years, the rise of deep learning methods has significantly driven technological innovation in the field of antibody design. Breakthroughs in protein structure prediction, exemplified by AlphaFold, have spurred the development of a series of models specifically designed for antibody design. For instance, RefineGNN achieved for the first time an autoregressive co-design of CDR sequences and structures; HSRN constructed a hierarchical optimization framework to integrate design and docking processes; DiffAb and AbDiffuser improved the quality of structure generation based on diffusion models; and MEAN and its improved version dyMEAN further optimized computational efficiency through all-atom modeling. Furthermore, ABGNN and HTP have explored the technical potential of sequence pre-training and hierarchical pre-training, respectively, opening up new directions for the development of antibody design methods.

[0007] Despite significant progress made by deep learning methods in antibody design, existing technologies still face several key challenges that limit their practical performance in downstream tasks.

[0008] First, the prediction accuracy of highly variable regions is insufficient, especially for intermediate amino acid residues far from the framework region in long CDR regions. Due to the difficulty in obtaining sufficient contextual structural information, the model's prediction accuracy for their coordinates and types is not high. This problem is common in different secondary structures such as helices, folds, and loops, and has become a major factor affecting antibody binding specificity.

[0009] Secondly, the quality control mechanism for training data is not yet perfect. Existing methods mostly use simple rules for screening and lack a systematic quantitative evaluation system for the rationality of antibody structures. Furthermore, antibody-antigen complex databases often contain data with structural deviations or abnormal physical properties, which directly limits the generalization ability of the model and reduces its predictive performance for unknown antigens, making it difficult to meet the requirements of industrial-grade platforms such as MoleculeOS for the accuracy of structural and physical properties.

[0010] Furthermore, the utilization of prior structural information remains insufficient. Although models such as AbDiffuser have attempted to introduce family-specific priors, atomic-level structural constraints, such as bond lengths, bond angles, and dihedral angles, have not yet been systematically integrated into the model evaluation and training process. Existing evaluations mostly rely on computational simulations or limited experimental verification, lacking systematic indicators that cover the entire structural hierarchy and adapt to various secondary structures. At the same time, they have failed to effectively integrate prior information throughout the entire process of model initialization and training optimization.

[0011] Therefore, constructing a quantitative evaluation system based on structural priors and improving the structural prediction accuracy of antibody variable regions, especially CDR regions, through high-quality data screening and model fine-tuning strategies has become an urgent technical problem to be solved in this field. Summary of the Invention

[0012] This invention addresses the aforementioned problems in the prior art by providing an antibody variable region structure prediction method based on structural prior assessment. This method solves the problems of low prediction accuracy of antibody CDR region structure, imperfect quality control of training data, and insufficient utilization of structural prior information, thereby effectively improving the accuracy of antibody variable region structure prediction.

[0013] To achieve the above objectives, this invention proposes a method for predicting the structure of antibody variable regions based on prior structural assessment, comprising: S1. Screening for complexes containing intact antigen and antibody H / L chains from PDB, retaining high-resolution, high-integrity samples, and extracting atomic coordinate data after structural processing; S2. Based on the extracted raw data, calculate the molecular structure geometric feature data, including bond length, bond angle, dihedral angle, and interatomic distance, as well as the spherical polar coordinates of adjacent residues in the relative spherical polar coordinate system; S3. The calculated data are used to calculate the cumulative mean square error loss and interval deviation loss as the structural quality assessment score based on the physical prior and the mapping relationship between dihedral angle and interatomic distance. The physical prior includes the standard values ​​of bond length and bond angle, and the range of spherical polar coordinates of adjacent residues relative to the spherical polar coordinate system. S4. Based on the structure quality assessment score, the antibody structure data for training is screened, and low-quality structure data is removed. S5. Receive antigen data, antibody framework data and high-quality dataset, construct antigen-antibody property heterogeneity diagram and initialize complex diagram; S6. In the model initialization stage, the antibody structure calculation data and the spherical polar coordinates of adjacent residues in the relative spherical polar coordinate system are embedded into the node features as enhanced feature representations. S7. Load the RA-EGN encoder to iteratively process the complex map, output the CDR sequence and 3D coordinates, and assemble them into a complete antibody structure; S8. During the model training process, a structural prior loss function is introduced to fine-tune the model, and the optimal antibody sequence and structure are selected and output through multi-index evaluation.

[0014] Preferably, in S1, the screening and structural processing specifically includes: S11. Screening for complex structures containing complete antigen chains, antibody heavy chains (H) and light chains (L) from the PDB and SAbDab databases. Screening criteria include resolution, residue integrity, and antigen type-related requirements. Irrelevant samples are excluded. S12. Use specialized tools to remove water molecules, ions, and other non-protein heteroatoms from the structure. Combine specialized tools to identify the H and L chains of the antibody, analyze intermolecular interactions to confirm the antigen chain, unify the identification of the three chains, and record their mapping relationship with the original identification.

[0015] Preferably, in S2, the bond length includes bond length types with different atomic compositions, the bond angle includes bond angle types with different atomic compositions, and the dihedral angle includes dihedral angle types with different atomic compositions and some values ​​fixed while others are variable; the formula for calculating the radial distance between adjacent residues in the spherical polar coordinate system is: ; Where (x',y',z') are the three-dimensional coordinates in the target atom space coordinate system; The formula for calculating the polar angle is: ; The azimuth angle uses the four-quadrant arctangent function, and its calculation formula is: .

[0016] Preferably, in S2, the parameters of the adjacent residues relative to the spherical polar coordinate system include radial distance, polar angle, and azimuth angle. The radial distance refers to the straight-line distance from the target atom to the origin. The polar angle refers to the angle between the target atom and the polar axis z-axis to describe the degree of deviation in the vertical direction. The azimuth angle refers to the angle between the target atom and the x-axis on the xy-plane projection to describe the horizontal orientation. Each parameter has a corresponding calculation method and value range. The adjacent residues relative to the spherical polar coordinate system includes multiple construction methods. Different construction methods correspond to different coordinate system poles, X-axis directions, and atoms to be described. Moreover, the radial distance, polar angle range, and azimuth angle range under each construction method have corresponding value requirements.

[0017] Preferably, in S3, the mapping relationship between the dihedral angle and the interatomic distance includes two types. One type is the mapping relationship between a specific dihedral angle ψ and the corresponding N0-N1 interatomic spacing, expressed as: ; Another type is a specific dihedral angle. dihedral The mapping relationship with the corresponding C0-C1 atomic spacing is expressed in the form of: .

[0018] Preferably, in S3, the calculation of the structural quality assessment score specifically includes: Cumulative mean square error loss calculation: Calculate the observed value y based on the geometric characteristics of bond length and bond angle. obs Compared with the standard value y std The mean square error is given by the formula: ; In the formula, N is the number of features, and the standard value is derived from the statistical average value of the PDB database; Interval Deviation Loss Calculation: For features with range constraints, a reasonable interval [y] is set. min ,ymax When the observed value y obs Calculate the loss value when the value exceeds the range: ; Mapping relationship loss calculation: based on dihedral angle θ dihedral The mapping relationship between f(θ) and interatomic spacing d dihedral =d, calculate the observation distance d obs Distance d from the prediction pred Mean square error: ; In the formula, K is the number of mapping pairs; Overall Score Calculation: The final structural quality assessment score is obtained through weighted summation. Weighting coefficients are used to balance the contributions of different losses. The score is negatively correlated with structural quality. The final structural quality assessment score is: Score = α·MSE cumulative +β·L interval +γ·L mapping ; In the formula, α, β, and γ are weighting coefficients.

[0019] Preferably, in S4, the screening process for the antibody structure data used for training is as follows: Based on the distribution characteristics of structural quality assessment scores and experimental requirements, a preset quality threshold is set. A comprehensive quality score is calculated for the antibody structural conformation variants corresponding to each PDB file. Low-quality samples with scores higher than the threshold are removed. The focus is on removing abnormal structures whose geometric features deviate significantly from the physical prior. The optimal conformation that meets the requirements is retained to construct a high-quality training dataset.

[0020] Preferably, in S6, the construction process of the enhanced feature representation is as follows: S61. Through distance encoding and angle encoding modules, bond length, bond angle and interatomic spacing are converted into numerical features; S62. The statistical constraints represented by the polar angle and azimuth angle of adjacent residues in the spherical coordinate system are embedded into the position encoding module to capture the spatial distribution pattern. S63. The above features are concatenated with the outputs of the type embedding and direction encoding modules to generate an enhanced node feature representation.

[0021] Preferably, in S8, the multi-index evaluation includes an index for assessing the accuracy of sequence generation and an index for measuring the quality of structure prediction; the structure prior loss function integrates bond length and bond angle deviation loss, dihedral angle distribution error, and spatial constraint loss in spherical coordinates, and together with the cross-entropy loss of sequence generation and the Huber loss of structure generation, constitutes a multi-task objective function, which jointly optimizes the model parameters through the backpropagation algorithm.

[0022] Preferably, in S5, the constructed antigen-antibody property heterogeneous graph has node features that include residue type embedding, position encoding, bond angle, dihedral angle encoding, distance encoding, orientation encoding, and antibody structure calculation data embedding, and edge features that include relation type encoding, relative position encoding, spatial distance / orientation encoding, and orientation quaternion encoding; in S7, the loaded RA-EGN encoder processes the complex graph through multiple rounds of iteration. The iteration process includes message aggregation, edge relation update, node and edge feature update, and node coordinate update, and finally outputs the CDR sequence and 3D coordinates.

[0023] Therefore, this invention proposes a method for predicting the structure of antibody variable regions based on prior structural assessment, which has the following advantages: (1) By constructing a quality assessment system based on structural geometry priors, training data is screened and low-quality samples such as abnormal bond lengths and bond angles and out-of-range spherical polar coordinates are eliminated to ensure the reliability of training data and solve the problem of imperfect quality control of existing training data.

[0024] (2) Incorporate structural priors such as the relative polar coordinates of adjacent residues into feature embedding and model training, and combine the RA-EGN encoder to collaboratively generate CDR sequences and 3D coordinates, thereby improving the accuracy of antibody CDR region structure prediction and overcoming the shortcomings of low prediction accuracy in existing models.

[0025] (3) Introduce a structural prior loss function fine-tuning model, screen the optimal results with multiple indicators, make full use of atomic-level structural constraint information, provide technical support for the design and optimization of therapeutic antibodies, and help the research and development of therapeutic antibodies for related diseases.

[0026] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0027] Figure 1 This is a flowchart of the antibody variable region structure prediction method based on prior structural assessment according to the present invention; Figure 2 This is a schematic diagram of the antibody-antigen complex data preprocessing flow of the antibody variable region structure prediction method based on structure prior assessment of the present invention. Figure 3 This is a flowchart of the antibody-antigen complex data structure quality assessment score calculation method based on the antibody variable region structure prediction method based on structure prior assessment of the present invention. Figure 4 This is a flowchart of the antibody-antigen complex data quality screening method based on the antibody variable region structure prediction method of the present invention. Figure 5 This is the model training process for the antibody variable region structure prediction method based on prior structural evaluation, as described in this invention. Figure 1 ; Figure 6 This is the model training process for the antibody variable region structure prediction method based on prior structural evaluation, as described in this invention. Figure 2 . Detailed Implementation

[0028] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0029] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0030] like Figures 1-6 As shown, an embodiment of the present invention provides an antibody variable region structure prediction method based on structural prior assessment. This embodiment is based on structural prior assessment technology and achieves accurate prediction of antibody variable region structure through steps such as data screening, feature calculation, model construction and training optimization.

[0031] The specific implementation steps are as follows: S1. Data Preprocessing and High-Quality Sample Screening: Database screening involves selecting complex structure samples from the PDB database that contain complete antigen chains, antibody heavy chains (H chains), and light chains (L chains).

[0032] The selection criteria are as follows: Resolution: X-ray diffraction sample resolution is no less than 3 Å, and cryo-electron microscopy sample resolution is no less than 4 Å; Residue integrity: The residue integrity of the H chain, L chain, and antigen chain must all reach more than 90%; Antigen type: Only samples with protein antigens are retained, while irrelevant samples such as polysaccharides and small molecule haptens are removed.

[0033] Structural treatment: The PyMOL tool was used to remove water molecules, ions, and other non-protein heteroatoms from the screened sample structure; the PyIgClassify tool was used to identify the H and L chains of the antibody, and the PDBePISA tool was used to analyze intermolecular interactions, identifying polypeptide chains that formed at least 5 hydrogen bonds with the antibody as antigen chains; the identifiers of the H chain, L chain, and antigen chain were standardized, and their mapping relationship with the chain identifiers in the original PDB file was recorded; the atomic coordinate data of the processed complex structure were extracted for subsequent geometric feature calculations.

[0034] S2. Molecular structural geometric features and spherical polar coordinate calculation: Calculation of molecular structure geometric features: Based on the atomic coordinate data extracted by S1, the geometric features of bond length, bond angle, dihedral angle, and interatomic distance were calculated. The atomic composition and prior reference information for each feature are shown in Table 1 below: Table 1. Prior values ​​of protein backbone bond length, bond angle, and dihedral angle geometric parameters.

[0035] Calculation of relative polar coordinates of adjacent residues: A relative spherical polar coordinate system of adjacent residues is constructed, and the spherical polar coordinates of the target atom are calculated. The specific calculation method and parameter range are shown in Table 2. Table 2. Definition and Calculation Method of Spherical Polar Coordinate System Parameters

[0036] The construction method and parameter range of the relative spherical polar coordinate system of adjacent residues are shown in Table 3: Table 3. Statistics on the range of parameters of adjacent residues relative to the spherical polar coordinate system.

[0037] S3. Calculation of Structural Quality Assessment Score: Based on physical priors, namely the standard values ​​of bond lengths and bond angles, the range of spherical polar coordinates, and the mapping relationship between dihedral angles and interatomic spacing, the structural quality assessment score is calculated, specifically including the following four types of loss calculations: Cumulative mean square error loss calculation: Calculate the observed value y based on the geometric characteristics of bond length and bond angle. obs Compared with the standard value y std The mean square error is given by the formula: ; In the formula, N is the number of features, and the standard value is derived from the statistical average value of the PDB database; Interval Deviation Loss Calculation: For features with range constraints, a reasonable interval [y] is set. min ,y max When the observed value y obs Calculate the loss value when the value exceeds the range: ; Mapping relationship loss calculation: based on dihedral angle θ dihedral The mapping relationship between f(θ) and interatomic spacing d dihedral =d, calculate the observation distance d obs Distance d from the prediction pred Mean square error: ; In the formula, K is the number of mapping pairs; The mapping relationships include two types: One type is the mapping relationship between a specific dihedral angle ψ and the corresponding N0-N1 atomic spacing, expressed as: ; Another type is a specific dihedral angle. dihedral The mapping relationship with the corresponding C0-C1 atomic spacing is expressed in the form of: .

[0038] Overall Score Calculation: The final structural quality assessment score is obtained through weighted summation. Weighting coefficients are used to balance the contributions of different losses. The score is negatively correlated with structural quality. The final structural quality assessment score is: Score = α·MSE cumulative +β·L interval +γ·L mapping ; In the formula, α, β, and γ are weighting coefficients.

[0039] S4. Training data selection: Based on the structural quality assessment score distribution characteristics calculated in step S3 and experimental requirements, a preset quality threshold is set. For each antibody structural conformation variant corresponding to a PDB file, a comprehensive quality score is calculated, and low-quality samples with scores higher than the threshold are removed. The following abnormal structures are specifically eliminated: Structures whose bond lengths and bond angles exceed ±5% of the prior constant values ​​in step S2; structures whose spherical polar coordinates exceed the corresponding range in step S2; structures whose absolute error in the mapping relationship between dihedral angles and interatomic spacing is greater than 0.2 Å; Retain the Top K optimal conformations that meet the requirements to construct a high-quality training dataset.

[0040] S5. Construction of antigen-antibody property isomer map and initialization of complex map: Input data preparation: Receive antigen data, antibody framework (FR) data, and the high-quality dataset obtained in step S4; Attribute heterogeneity graph construction: Node features: covering residue type embedding, position encoding, bond angle / dihedral angle encoding, distance encoding, orientation encoding, and antibody structure computational data embedding; Edge features include relation type encoding, relative position encoding, spatial distance / direction encoding, and orientation quaternion encoding; Complex diagram initialization: Remove the CDR regions HCDR1-3 to be generated in the antibody heavy chain, and use the HDOCK tool to perform molecular docking of the remaining antibody framework with the antigen to obtain the initial antigen-antibody complex structure; Based on the coordinates of adjacent residues before and after the CDR region, initialize the 3D coordinates of N, CA, C, and O atoms of the CDR region main chain in a uniform distribution to form an incomplete complex diagram.

[0041] S6. Model Initialization and Enhanced Feature Embedding: In the initialization phase of the RA-EGN-based graph neural network model, enhanced feature representations are constructed, with the following specific steps: Distance and angle encoding: The distance encoding module maps bond length and interatomic distance to a normalized function in the range [0,1]. The angle encoding module converts bond angle and dihedral angle into sine or cosine values, and converts the bond length, bond angle and interatomic distance calculated in step S2 into numerical features. Spherical polar coordinate embedding: The polar angle and azimuth angle of adjacent residues relative to the spherical polar coordinate system are embedded into the position encoding module, and the spatial distribution pattern is captured by trigonometric function transformation; Feature concatenation: The distance / angle encoded features, spherical polar coordinate embedding features, and type embeddings, such as type identifiers for "antigen residues" and "antibody residues", and the output of the direction encoding, i.e. the normalized value of the inter-atomic direction vector, are concatenated to generate an enhanced node feature representation for model input.

[0042] S7 and RA-EGN encoder processing and complete antibody structure generation: RA-EGN Encoder Loading and Iterative Processing: Load the pre-trained Relation-Aware Isomorphic Graph Network (RA-EGN) encoder and perform multiple rounds of iterative processing on the complex graph initialized in step S5. Each round of iteration includes: Message Aggregation: Calculate the message passing weights of neighboring nodes based on edge features and aggregate neighbor node information; Edge Relationship Update: Dynamically adjust the relationship type encoding and spatial distance or direction encoding of edges according to changes in node features; Node or Edge Feature Update: Update node features and edge features through a Multilayer Perceptron (MLP); Node Coordinate Update: Adjust the 3D coordinates of nodes (atoms) based on isomorphic transformation rules. CDR Sequence and 3D Coordinate Output: After the iteration, the amino acid sequence of the CDR region is predicted by the sequence generation head of the model, and the precise 3D coordinates of the atoms in the CDR region are output by the structure generation head. Complete antibody structure assembly: The generated CDR sequence and 3D coordinates are assembled with the antibody framework in step S5 to form an antibody-antigen complex structure containing the complete H chain, L chain and antigen.

[0043] S8 Model Fine-tuning and Optimal Result Selection: Introduction of the structural prior loss function: During model training, a structural prior loss function is introduced, which, together with the cross-entropy loss from sequence generation and the Huber loss from structure generation, constitutes the multi-task objective function. Structural prior losses: integrating bond length and bond angle deviation losses, dihedral angle distribution errors, and spherical polar coordinate constraint losses; Cross-entropy loss: used to optimize the accuracy of CDR sequence prediction by calculating the cross-entropy between the predicted amino acid type and the actual type; Huber loss: Used to optimize the 3D coordinate prediction accuracy of CDR structure, the Huber loss between predicted coordinates and true coordinates, reducing the impact of outliers; Model parameter optimization: The parameters of the RA-EGN encoder and generator head are iteratively updated based on the multi-task objective function through backpropagation algorithm until the model converges; Multi-index evaluation and optimal result selection: The antibody sequences and structures output by the model are evaluated using the following indicators to select the optimal results; Sequence accuracy metric: Amino acid recovery rate (AAR), which is the proportion of identical amino acids in the predicted CDR sequence and the true sequence; Structural quality metrics: Root Mean Square Deviation (RMSD), the deviation of atomic coordinates between the predicted CDR structure and the actual structure, the smaller the better; Template Modeling Score (TM-score), which measures structural similarity, the closer to 1 the better; The loss function based on antibody structure prior is associated with the above evaluation indicators. Backpropagation of the loss function enables model fine-tuning and training correction, reducing structure prediction bias. Finally, the optimal complex corresponding CDR sequence and complete antibody structure are output, supporting the adaptation verification of long sequence CDR and unknown epitope scenarios, and outputting the optimal antibody sequence and complete antibody-antigen complex structure.

[0044] Based on the above method, an antibody variable region structure prediction device based on prior quality assessment of antibody structure is constructed. This device includes the following modules: Antibody structure data preprocessing and quality screening module: Performs steps S1-S4, including PDB / SAbDab data screening, structure decontamination and chain identification, geometric feature and spherical polar coordinate calculation, structure quality assessment and low-quality data removal, and outputs a high-quality training dataset; Attribute Heterogeneous Graph Construction and Encoding Module: Executes steps S5-S6, receives antigen and antibody frameworks and high-quality data, constructs antigen-antibody attribute heterogeneous graphs, generates enhanced node or edge features containing structural prior information, and outputs encoded complex graphs. Complex diagram initialization module: Executes the complex diagram initialization operation in step S5, removes the region to be generated CDR, completes antigen-antibody docking, initializes CDR atom coordinates, and outputs an incomplete complex diagram; Relationship-aware sequence-structure co-generation module: Executes the operation in step S7, loads the RA-EGN encoder to iteratively process the complex map, outputs the CDR sequence and 3D coordinates, and assembles the complete antibody structure; Multi-dimensional performance evaluation and screening module: Execute the operation in step S8, introduce the structural prior loss function to fine-tune the model, screen the optimal antibody sequence and structure through multi-index evaluation, and output the final result.

[0045] Furthermore, the technical solution of the present invention can be implemented through the following three types of carriers: The computer device includes a memory, a processor, and a computer program stored in the memory. When the processor executes the program, it implements all the operations in steps S1-S8 of this embodiment. The memory can be a solid-state drive (SSD) or a hard disk drive (HDD), and the processor can be a CPU, supporting GPU-accelerated model training.

[0046] The computer-readable storage medium is a non-transitory storage medium that stores a computer program. When the program is executed by a processor, it implements the method flow of steps S1-S8 in this embodiment.

[0047] The computer program product contains a computer program that is stored in a non-transitory storage medium. When the program is executed, it drives the computer to complete the entire process of predicting the structure of the antibody variable region, which meets the protection requirements of software inventions.

[0048] Therefore, this invention provides a method for predicting the structure of antibody variable regions based on structural prior assessment. It involves screening high-quality antibody-antigen complex data from databases such as PDB, calculating the molecular structure's geometric features and the polar coordinates of adjacent residues, and using a physical prior to construct a quality assessment system for screening training data. Then, it constructs an antigen-antibody property heterogeneity graph, embeds structural prior features into the model initialization, and uses an RA-EGN encoder to collaboratively generate CDR sequences and 3D coordinates. A structural prior loss function is introduced to fine-tune the model, and multiple indicators are used to screen for the optimal result. This effectively solves the problems of low prediction accuracy of antibody CDR regions, inadequate quality control of training data, and insufficient utilization of structural priors in existing methods, improving the accuracy of antibody variable region structure prediction and providing reliable technical support for the design and optimization of therapeutic antibodies.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the structure of antibody variable regions based on prior structural assessment, characterized in that, include: S1. Screening for complexes containing intact antigen and antibody H / L chains from PDB, retaining high-resolution, high-integrity samples, and extracting atomic coordinate data after structural processing; S2. Based on the extracted raw data, calculate the molecular structure geometric feature data, including bond length, bond angle, dihedral angle, and interatomic distance, as well as the spherical polar coordinates of adjacent residues in the relative spherical polar coordinate system; S3. The calculated data are used to calculate the cumulative mean square error loss and interval deviation loss as the structural quality assessment score based on the physical prior and the mapping relationship between dihedral angle and interatomic distance. The physical prior includes the standard values ​​of bond length and bond angle, and the range of spherical polar coordinates of adjacent residues relative to the spherical polar coordinate system. S4. Based on the structure quality assessment score, the antibody structure data for training is screened, and low-quality structure data is removed. S5. Receive antigen data, antibody framework data and high-quality dataset, construct antigen-antibody property heterogeneity diagram and initialize complex diagram; S6. In the model initialization stage, the antibody structure calculation data and the spherical polar coordinates of adjacent residues in the relative spherical polar coordinate system are embedded into the node features as enhanced feature representations. S7. Load the RA-EGN encoder to iteratively process the complex map, output the CDR sequence and 3D coordinates, and assemble them into a complete antibody structure; S8. During the model training process, a structural prior loss function is introduced to fine-tune the model, and the optimal antibody sequence and structure are selected and output through multi-index evaluation.

2. The antibody variable region structure prediction method based on prior structural assessment according to claim 1, characterized in that, In S1, the screening and structural processing specifically include: S11. Screening for complex structures containing complete antigen chains, antibody heavy chains (H) and light chains (L) from the PDB and SAbDab databases. Screening criteria include resolution, residue integrity, and antigen type-related requirements. Irrelevant samples are excluded. S12. Use specialized tools to remove water molecules, ions, and other non-protein heteroatoms from the structure. Combine specialized tools to identify the H and L chains of the antibody, analyze intermolecular interactions to confirm the antigen chain, unify the identification of the three chains, and record their mapping relationship with the original identification.

3. The antibody variable region structure prediction method based on structural prior assessment according to claim 1, characterized in that, In S2, the bond length includes bond length types with different atomic compositions, the bond angle includes bond angle types with different atomic compositions, and the dihedral angle includes dihedral angle types with different atomic compositions and some values ​​fixed while others are variable; the formula for calculating the radial distance between adjacent residues in the spherical polar coordinate system is: ； Where (x',y',z') are the three-dimensional coordinates in the target atom space coordinate system; The formula for calculating the polar angle is: ； The azimuth angle uses the four-quadrant arctangent function, and its calculation formula is: 。 4. The antibody variable region structure prediction method based on structural prior assessment according to claim 1, characterized in that, In S2, the parameters of the adjacent residues relative to the spherical polar coordinate system include radial distance, polar angle, and azimuth angle. Radial distance refers to the straight-line distance from the target atom to the origin. Polar angle refers to the angle between the target atom and the polar axis z-axis to describe the degree of deviation in the vertical direction. Azimuth angle refers to the angle between the target atom and the x-axis on the xy-plane projection to describe the horizontal orientation. Each parameter has a corresponding calculation method and value range. The adjacent residues relative to the spherical polar coordinate system includes multiple construction methods. Different construction methods correspond to different coordinate system poles, X-axis directions, and atoms to be described. Moreover, the radial distance, polar angle range, and azimuth angle range under each construction method have corresponding value requirements.

5. The antibody variable region structure prediction method based on prior structural assessment according to claim 1, characterized in that, In S3, the mapping relationship between the dihedral angle and the interatomic distance includes two types of expressions. One type is the mapping relationship between a specific dihedral angle ψ and the corresponding N0-N1 interatomic spacing, expressed as follows: ； Another type is a specific dihedral angle. dihedral The mapping relationship with the corresponding C0-C1 atomic spacing is expressed in the form of: 。 6. The antibody variable region structure prediction method based on structural prior assessment according to claim 1, characterized in that, In S3, the calculation of the structural quality assessment score specifically includes: Cumulative mean square error loss calculation: Calculate the observed value y based on the geometric characteristics of bond length and bond angle. obs Compared with the standard value y std The mean square error is given by the formula: ； In the formula, N is the number of features, and the standard value is derived from the statistical average value of the PDB database; Interval Deviation Loss Calculation: For features with range constraints, a reasonable interval [y] is set. min ,y max When the observed value y obs Calculate the loss value when the value exceeds the range: ； Mapping relationship loss calculation: based on dihedral angle θ dihedral The mapping relationship between f(θ) and interatomic spacing d dihedral =d, calculate the observation distance d obs Distance d from the prediction pred Mean square error: ； In the formula, K is the number of mapping pairs; Overall Score Calculation: The final structural quality assessment score is obtained through weighted summation. Weighting coefficients are used to balance the contributions of different losses. The score is negatively correlated with structural quality. The final structural quality assessment score is: Score=α·MSE cumulative +β·L interval +γ·L mapping ; In the formula, α, β, and γ are weighting coefficients.

7. The antibody variable region structure prediction method based on prior structural assessment according to claim 1, characterized in that, In S4, the screening process for the antibody structure data used for training is as follows: Based on the distribution characteristics of structural quality assessment scores and experimental requirements, a preset quality threshold is set. A comprehensive quality score is calculated for the antibody structural conformation variants corresponding to each PDB file. Low-quality samples with scores higher than the threshold are removed. The focus is on removing abnormal structures whose geometric features deviate significantly from the physical prior. The optimal conformation that meets the requirements is retained to construct a high-quality training dataset.

8. The antibody variable region structure prediction method based on structural prior assessment according to claim 1, characterized in that, In S6, the construction process of the enhanced feature representation is as follows: S61. Through distance encoding and angle encoding modules, bond length, bond angle and interatomic spacing are converted into numerical features; S62. The statistical constraints represented by the polar angle and azimuth angle of adjacent residues in the spherical coordinate system are embedded into the position encoding module to capture the spatial distribution pattern. S63. The above features are concatenated with the outputs of the type embedding and direction encoding modules to generate an enhanced node feature representation.

9. The antibody variable region structure prediction method based on structural prior assessment according to claim 1, characterized in that, In S8, the multi-index evaluation includes an index for assessing the accuracy of sequence generation and an index for measuring the quality of structure prediction; the structure prior loss function integrates bond length and bond angle deviation loss, dihedral angle distribution error, and spatial constraint loss in spherical coordinates, and together with the cross-entropy loss of sequence generation and the Huber loss of structure generation, it constitutes a multi-task objective function, which jointly optimizes the model parameters through the backpropagation algorithm.

10. The antibody variable region structure prediction method based on prior structural assessment according to claim 1, characterized in that, In S5, the constructed antigen-antibody property heterogeneity graph has node features that include residue type embedding, position encoding, bond angle, dihedral angle encoding, distance encoding, orientation encoding, and antibody structure calculation data embedding. Edge features include relation type encoding, relative position encoding, spatial distance / orientation encoding, and orientation quaternion encoding. In S7, the loaded RA-EGN encoder processes the complex graph through multiple rounds of iteration. The iteration process includes message aggregation, edge relation update, node and edge feature update, and node coordinate update, and finally outputs the CDR sequence and 3D coordinates.

Images