An Antibody Design Method Based on Flow Matching and Graph Neural Networks
By combining flow matching with graph neural networks, the coupling of local and global structures in antibody design is achieved, solving the problem of inconsistent antibody generation in existing technologies and improving the structural accuracy and computational efficiency of antibody binding regions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- RENMIN UNIVERSITY OF CHINA
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-26
AI Technical Summary
Existing antibody design methods lack three-dimensional structural constraints during the generation process, resulting in inconsistent antibody conformations and insufficient antigen-binding interface quality, making it difficult to achieve efficient and accurate antibody design.
We employ a method based on flow matching and graph neural networks. By constructing an antibody graph structure and utilizing a global geometric message passing mechanism driven by antigen surface point clouds, we achieve local generation of antibody binding regions and global structure propagation. We then combine flow matching ODE and surface-enhanced message encoders to perform end-to-end antibody design.
This approach achieves coupling between local and global structures in the antibody design process, improves the structural consistency of the antibody binding region and the quality of the contact interface, reduces computational overhead, and forms an efficient antibody design scheme.
Smart Images
Figure CN122090916A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of computer science, and more specifically, to an antibody design method based on flow matching and graph neural networks. Background Technology
[0002] Antibodies, as important biomolecules, have wide applications in disease treatment, diagnosis, and immunoengineering. The specific binding of antibodies to antigens is mainly determined by complementarity-determining regions (CDRs) within the variable region. Among these, CDR-H3 (paratope), due to its highly diverse structure, plays a crucial role in antigen recognition for most antibodies. Therefore, the ability to efficiently design antibody sequences and three-dimensional structures based on specific antigens is an important research direction in the field of antibody engineering.
[0003] With the development of deep learning technology, an increasing number of studies are attempting to utilize data-driven methods for antibody design. One type of method relies on sequence-level generation models, which generate candidate sequences by learning large-scale antibody sequence distributions. However, these methods lack three-dimensional structural constraints, making it difficult to guarantee the conformational feasibility and binding interface quality of the generated antibody. Another type of method introduces three-dimensional structural information, employing structure prediction models, graph neural networks, or diffusion models to locally generate or optimize the CDR region. However, these methods typically focus only on local fragments (such as CDR-H3) or employ a step-by-step process, with each functional module operating independently. This results in a lack of consistency between the local structure and the overall conformation, making it difficult to guarantee the overall stability and accuracy of the generated antibody at the atomic scale.
[0004] In recent years, some end-to-end antibody generation frameworks have emerged, attempting to simultaneously model sequences and three-dimensional structures. However, existing methods generally have two shortcomings: first, insufficient coupling between the generation process and the transfer of structural information, resulting in limited information interaction between paratope generation and the overall antibody conformation; second, insufficient utilization of the fine geometric features of the antigen surface, making it difficult for the generated antibody-antigen binding interface to achieve high structural quality or affinity performance.
[0005] Therefore, it is necessary to propose an antibody design method that can strengthen the coupling between local and global structures, make full use of antigen surface geometric information, and support end-to-end generation at the atomic level, so as to effectively overcome the limitations of existing technologies. Summary of the Invention
[0006] The purpose of this disclosure is to provide an antibody design method based on flow matching and graph neural networks, which aims to solve the problem of generating high-quality antibody binding region sequences and all-atom three-dimensional structures under the constraint of antigen surface geometry information.
[0007] In general, this paper presents an antibody design method based on flow matching and graph neural networks, which is implemented through the following steps: Step 1: Input parsing and initial antibody structure construction. For the input three-dimensional structure of the antigen epitope and the antibody sequence missing the paratope, the antigen-antibody input information is parsed and converted, the antibody nodes and coordinate features are initialized, and the antibody graph structure is constructed. Step 2: Model the surface of the antigen epitope, generate a point cloud of the antigen epitope surface, assign point clouds to epitope residues, and sample and fill in fixed surface vertices; Step 3: Local generation of the binding region based on flow matching. Using sequence, coordinates, and surface vertices as conditions, flow matching is used to solve the ODE to generate local 3D coordinates. Step 4: The surface enhancement message encoder is globally propagated to construct surface enhancement messages, jointly update node and coordinate features, and construct an integrated flow matching velocity field network; Step 5: Sequence-structure joint convergence generation. The paratope structure is updated in multiple steps through the velocity field network. The coordinates converge to the target conformation, and the node features converge to the residue sequence. Finally, the all-atom antibody 3D structure and paratope residue sequence are output.
[0008] The specific implementation method of the input reading and initial antibody structure construction process is as follows: Step 1-1: Input data reading and parsing. Read the three-dimensional structure of the antigen epitope and obtain its residues and corresponding coordinate information; at the same time, read the known amino acid sequence and IMGT number of the antibody, and extract the three-dimensional coordinates of the known main chain atoms of the antibody, parse and organize them into the structured input required for subsequent modeling. Steps 1-2: Construction of the initial antibody structure representation, including the antibody... Represented as a graph structure ,in For a set of residue nodes, This is a set of edges connecting residues. Specifically, each antibody residue is treated as a node. initialize node features This feature is encoded by the residue type and its three-dimensional coordinates of all atoms. The basic information is used to construct the set; then, the edge set is constructed based on the spatial distance relationship between the residues. If the initial main chain coordinate distance between two residues does not exceed a preset threshold. If the two are connected, then an edge is formed. Finally, the initial structure of the antibody is obtained: .
[0009] The specific implementation method of the antigen epitope surface modeling process is as follows: Step 2-1: Generate antigen surface point cloud. Generate surface point cloud data based on the three-dimensional atomic coordinates of the antigen molecule, and output a set of vertex coordinates located on the molecular surface. Step 2-2: Divide the surface point cloud by residue. Based on the correspondence between surface points and their corresponding atoms, assign the generated surface vertices to their corresponding antigen residues to obtain the surface vertex set for each residue: , where each surface point The three-dimensional spatial coordinates are represented as , This represents the actual number of surface vertices contained in the residue. Steps 2-3: Fixed-number vertex sampling and padding, performing a fixed number of samples for each residue: when At that time, from the set A fixed number of random samples One vertex; when When the time is right, all existing vertices are retained, and the remaining parts are filled with zero vectors; The final standardized set of surface vertices is obtained as follows: .
[0010] The specific implementation of the local generation process of the binding region based on flow matching is as follows: Step 3-1: Construct the conditional input for flow matching generation. Use the 3D coordinates of the paratope region as the object for generating the flow matching ODE. Use its initial node features, coordinates, and the set of antigen surface vertices obtained in Step 2 as conditions for each binding region node. ,structure As its moment The initial state; Step 3-2: Generate the local structure of the paratope using flow matching, treating the paratope coordinates as continuous variables that evolve over time, and introduce ODE: ,in A learnable flow-matched velocity field.
[0011] The specific implementation process of global structure propagation based on surface-enhanced message encoder is as follows: Step 4-1: Construct a surface enhancement message encoder. First, utilize the antigen surface vertex set generated in Step 2. Based on the spatial relationship between paratope residues and surface vertices, a fine geometric message oriented towards the paratope is constructed. The fine geometric message is based on continuous geometric difference and orientation-sensitive operators to realize the local structural transfer from the antigen surface to the paratope. message functions oriented towards three-dimensional coordinates The relationship between surface information and paratope coordinates is established using the following expression: ,in Represents coordinate difference. For learnable transformations; Step 4-2: Feature and coordinate update of the surface enhancement message encoder. Joint update of features and coordinates is performed in the antibody graph structure to propagate the local geometric influence of the antigen surface point cloud to the global structure. Specifically: for an edge in the antibody graph... First, the node features , The node edge messages are obtained by jointly encoding the surface-enhanced geometric messages between them. Subsequently, each node aggregates the edge messages of all neighboring nodes and inputs them into the feature update function to obtain new node features. ; A geometric update is constructed based on the relative coordinate difference between a node and its neighbors, and then added to the current node coordinates to obtain the new coordinates. Implement coordinate updates; Step 4-3: Integrate the surface augmentation message encoder with the velocity field network, directly incorporating the features and coordinate update mechanism of the surface augmentation message encoder into the velocity field network of the flow matching framework. This enables unified structure updates that combine local driving and global coordination.
[0012] The specific implementation process of the joint convergence generation of the sequence and structure is as follows: Velocity Field Network The structure of the paratope is continuously updated based on the current node features and 3D coordinates. Simultaneously, by leveraging a surface-enhanced message encoder and a variant of the isomorphic graph neural network integrated into the network, local geometric changes are propagated across the entire antibody, achieving synchronous adjustment of the overall structure. Over time, the antibody's atomic coordinates gradually converge to a stable conformation, and the node features also converge to the corresponding amino acid distribution, thus generating a consistent 3D antibody structure. and binding region sequence This completes the integrated generation of sequences and structures.
[0013] The technical effects to be achieved by the embodiments of the present invention are as follows: By constructing a flow-matching generation process in the binding region and utilizing a global geometric message passing mechanism driven by antigen surface point clouds, more accurate interface modeling and structure generation are achieved. At the same time, restricting flow-matching calculations to local regions effectively reduces the computational overhead of macromolecular structure generation. This invention maintains significant structural accuracy advantages while possessing acceptable inference costs and good scalability, thus forming an efficient, end-to-end antibody design scheme. Attached Figure Description
[0014] The above and other objects and features of this disclosure will become clearer from the following description taken in conjunction with the accompanying drawings.
[0015] Figure 1 This is a schematic diagram illustrating the architecture of an antibody design method based on flow matching and graph neural networks according to embodiments of the present disclosure; Figure 2 This is a schematic diagram illustrating the single-step update principle of a surface-enhanced flow-matched velocity field network according to an embodiment of the present disclosure. Detailed Implementation
[0016] The following detailed embodiments are provided to aid the reader in gaining a comprehensive understanding of the methods, apparatus, and / or systems described herein. However, various changes, modifications, and equivalents of the methods, apparatus, and / or systems described herein will become apparent upon understanding this disclosure. For example, the order of operations described herein is merely illustrative and is not limited to those orders set forth herein, but may be changed as will become clear upon understanding this disclosure, except for operations that must occur in a specific order. Furthermore, for clarity and conciseness, descriptions of features known in the art may be omitted.
[0017] The features described herein may be implemented in different forms and should not be construed as limited to the examples described herein. Rather, the examples described herein are provided only to illustrate some of the many feasible ways of implementing the methods, apparatus, and / or systems described herein, which will become clear upon understanding the disclosure of this application.
[0018] As used herein, the term “and / or” includes any one of the associated listed items and any combination of any two or more.
[0019] Although terms such as “first,” “second,” and “third” may be used herein to describe various components, assemblies, regions, layers, or parts, these components, assemblies, regions, layers, or parts should not be limited by these terms. Rather, these terms are used only to distinguish one component, assembly, region, layer, or part from another. Thus, without departing from the teaching of the examples described herein, the first component, first assembly, first region, first layer, or first part referred to as the first component, first assembly, first region, first layer, or first part may also be referred to as the second component, second assembly, second region, second layer, or second part.
[0020] In the specification, when an element (such as a layer, region, or substrate) is described as being "on" another element, "connected to," or "bonded to" another element, the element may be directly "on" another element, directly "connected to," or "bonded to" the other element, or one or more other elements may be present in between. Conversely, when an element is described as being "directly on" another element, "directly connected to," or "directly bonded to" another element, no other elements may be present in between.
[0021] The terminology used herein is for the purpose of describing various examples only and is not intended to limit disclosure. Unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. The terms “comprising,” “including,” and “having” indicate the presence of the described features, quantities, operations, components, elements, and / or combinations thereof, but do not preclude the presence or addition of one or more other features, quantities, operations, components, elements, and / or combinations thereof.
[0022] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains upon understanding this disclosure. Unless expressly defined herein, terms (such as those defined in a general dictionary) shall be interpreted as having a meaning consistent with their meaning in the context of the relevant field and in this disclosure, and shall not be interpreted in an idealized or overly formalistic manner.
[0023] Furthermore, in the description of the examples, detailed descriptions of well-known related structures or functions will be omitted when it is believed that such detailed descriptions would lead to a vague interpretation of this disclosure.
[0024] Figure 1 This is a schematic diagram illustrating an antibody design method based on flow matching and graph neural networks according to an embodiment of the present disclosure.
[0025] To achieve the aforementioned objectives, the present invention employs the following technical framework: Figure 1 As shown.
[0026] This invention proposes an antibody design method based on flow matching and surface structure enhancement graph neural networks. Even with known epitope structures and missing sequences in the antibody's critical binding region (a local region within the variable region, the paratope), this method can achieve the desired amino acid sequence for the antibody's critical binding region. Generate and restore the full atomic three-dimensional structure of the entire antibody. This invention fully utilizes the fine-grained geometric information of the antigen surface, generates binding region structures through local flow matching, and uses graph neural networks to propagate local structural information to the overall antibody structure, thereby significantly improving the overall consistency of the antibody structure and the quality of the contact interface. The overall process can be illustrated in... Figure 1 It consists of the following steps, which are executed sequentially and are interdependent, forming a complete end-to-end antibody design process.
[0027] Step 1: Input Reading and Initial Antibody Structure Construction Step 1-1: Input Data Reading and Parsing This step involves reading the three-dimensional structure of the antigen epitope, obtaining its residues and corresponding coordinate information; simultaneously, it reads the known amino acid sequence and IMGT number of the antibody, and extracts the three-dimensional coordinates of the known main chain atoms of the antibody. This data is then parsed and organized into the structured input required for subsequent modeling.
[0028] Step 1-2: Construction of Initial Antibody Structure Representation After completing the data parsing, the antibodies are represented as a graph structure. Each antibody residue serves as a node. initialize node features This feature is encoded by the residue type and its three-dimensional coordinates of all atoms. These basic information elements are used together to form the basis. Then, an edge set is constructed based on the spatial distance relationships between residues. If the initial main chain coordinate distance between two residues does not exceed a preset threshold. If the two are connected, then an edge is formed. By using the node set, edge set, and initializing node features and coordinates, the initial graph structure of the antibody is obtained: .
[0029] This graph structure provides a basic representation for subsequent integration of region generation and global structural information propagation.
[0030] Step 2: Antigen epitope surface modeling This step is used to generate a surface point cloud based on the three-dimensional structure of the antigen, and to construct a fixed-length set of surface vertices for each residue as conditional inputs for subsequent generation processes. Specifically, it includes the following steps: Step 2-1: Generate antigen surface point cloud Using the molecular surface computation tool MSMS, surface point cloud data is generated based on the three-dimensional atomic coordinates of the antigen molecule. The output is a set of vertex coordinates located on the molecular surface.
[0031] Step 2-2: Divide the surface point cloud according to residues Based on the correspondence between surface points and their corresponding atoms, the generated surface vertices are assigned to their corresponding antigen residues, resulting in a set of surface vertices for each residue: , where each surface point The three-dimensional spatial coordinates are represented as , This represents the actual number of surface vertices contained in the residue.
[0032] Steps 2-3: Fixed number of vertex sampling and padding To maintain a consistent dimension for subsequent model inputs, a fixed number of samples are taken for each residue: when At that time, from the set A fixed number of random samples One vertex; when When the time is right, all existing vertices are retained, and the remaining parts are filled with zero vectors.
[0033] This yields the standardized set of surface vertices: This set contains the antigen surface spatial information required for subsequent steps.
[0034] Step 3: Local generation of the binding region based on flow matching Step 3-1: Construct the conditional input for stream matching generation This step uses the 3D coordinates of the paratope region as the object for generating the flow-matched ODE, using its initial node features, coordinates, and the set of antigen surface vertices obtained in step two as conditions. For each binding region node... ,structure As its moment The initial state.
[0035] Step 3-2: Generate local paratope structure using flow matching Treating the paratope coordinates as a continuous variable that evolves over time, we introduce ODE. , in This is a learnable flow-matching velocity field. By numerically solving this equation, the paratope coordinates can gradually approximate the target distribution, yielding the initially generated local three-dimensional coordinates of the bonding region, providing basic geometric information for subsequent global structure propagation.
[0036] Step 4: Global Structure Propagation Based on Surface-Enhanced Message Encoder The core of this step is the Surface Multi-channel Encoder (SME). SME models the local geometric information of the antigen surface point cloud and the structural information between residues inside the antibody in a unified manner, and uses a coordinate-variable message passing mechanism to propagate this information throughout the entire antibody structure.
[0037] Step 4-1: Construction of Surface Enhancement Messages SME first utilizes the set of antigen surface vertices generated in step two. Based on the spatial relationship between paratope residues and surface vertices, a fine-grained geometric message (SME) is constructed oriented towards the paratope. This message, based on continuous geometric difference and orientation-sensitive operators, realizes the local structural transfer from the antigen surface to the paratope. SME utilizes a message function oriented towards three-dimensional coordinates. The relationship between surface information and paratope coordinates is established, and its expression is: , in Represents coordinate difference. It is a learnable transformation.
[0038] Step 4-2: Feature and Coordinate Update of Surface Augmentation Message Encoder After constructing the surface geometry-based message, SME performs a joint update of features and coordinates within the antibody graph structure to propagate the local geometric influence of the antigen surface point cloud to the global structure. For an edge in the antibody graph... SME first identifies node features , The node edge messages are obtained by jointly encoding the surface-enhanced geometric messages between them. Subsequently, each node aggregates the edge messages of all neighboring nodes and inputs them into the feature update function to obtain new node features. This allows the fine-grained surface geometry signals received near the paratope to gradually propagate into the overall antibody structure.
[0039] For coordinate updates, SME constructs a geometric update amount based on the relative coordinate difference between a node and its neighborhood, and then adds it to the current node coordinates to obtain the new coordinates. This process maintains strict rotation-translation isotropy and ensures that changes in coordinates are consistent with changes in features, guaranteeing that local geometric perturbations can generate a coordinated structural response across the entire antibody range.
[0040] After multiple layers of propagation, SME achieves unified modeling of the geometric information of the antigen surface, the structural dependencies within the antibody, and the characteristics of the paratope region within the same message passing framework.
[0041] Step 4-3: Integration of SME with the velocity field network The SME's features and coordinate update mechanism are directly incorporated into the velocity field network of the flow matching framework. It is used to calculate the time evolution of antibody structure during the generation process. It also integrates an equivariant graph neural network variant, AME (Atom Multi-channel Encoder), employing an adaptive multi-channel equivariant message passing mechanism: a geometric relation extractor calculates the Euclidean distance matrix of all atomic pairs between residues, and aggregates them into fixed-dimensional geometric features using learnable channel weights and attribute matrices; subsequently, a geometric message scaler scales the coordinate differences to generate equivariant coordinate updates. Specifically, during the update... At each time step Receive current antibody sequence and coordinate status The state update for the next time step is output through joint encoding of SME and AME. During information exchange, SME directly transmits surface-level geometric messages between epitopes and paratopes, while AME transmits them across the entire antibody map. The message passing process broadcasts the local updates of the paratope to the complete antibody structure. The alternating iterations of the two ensure that the flow matching drive of the local paratope is coordinated and unified with the global antibody structure, thus guaranteeing the overall geometric rationality and the generation of full atomic-level coordinates during the continuous generation process.
[0042] Step 5: Sequence-Structure Joint Convergence Generation Based on Velocity Field Velocity field network within the flow matching framework The structure of the paratope is continuously updated based on the current node features and 3D coordinates. Simultaneously, by leveraging SME and isomorphic graph neural network variants integrated into the network, local geometric changes are propagated across the entire antibody, achieving synchronous adjustment of the overall structure. Over time, the antibody's atomic coordinates gradually converge to a stable conformation, and the node features also converge to the corresponding amino acid distribution, thus generating a consistent 3D antibody structure. and binding region sequence This completes the integrated generation of sequences and structures.
[0043] After the binding region sequence decoding and 3D structure alignment are completed, the system's final output includes two parts: the atomic-level 3D structural coordinates of the complete antibody and the predicted amino acid sequence of the binding region residues. This output can be directly used for subsequent downstream antibody design and analysis, such as affinity screening, interface optimization, or experimental validation.
[0044] While some embodiments of this disclosure have been shown and described, those skilled in the art will understand that modifications may be made to these embodiments without departing from the principles and spirit of this disclosure, which are defined by the claims and their equivalents.
Claims
1. An antibody design method based on flow matching and graph neural networks, characterized in that, The following steps are required to achieve this: Step 1: Input parsing and initial antibody structure construction. For the input three-dimensional structure of the antigen epitope and the antibody sequence missing the paratope, the antigen-antibody input information is parsed and converted, the antibody nodes and coordinate features are initialized, and the antibody graph structure is constructed. Step 2: Model the surface of the antigen epitope, generate a point cloud of the antigen epitope surface, assign point clouds to epitope residues, and sample and fill in fixed surface vertices; Step 3: Local generation of the binding region based on flow matching. Using sequence, coordinates, and surface vertices as conditions, flow matching is used to solve the ODE to generate local 3D coordinates. Step 4: The surface enhancement message encoder is globally propagated to construct surface enhancement messages, jointly update node and coordinate features, and construct an integrated flow matching velocity field network; Step 5: Sequence-structure joint convergence generation. The paratope structure is updated in multiple steps through the velocity field network. The coordinates converge to the target conformation, and the node features converge to the residue sequence. Finally, the all-atom antibody 3D structure and paratope residue sequence are output.
2. The antibody design method based on flow matching and graph neural networks as described in claim 1, characterized in that, The specific implementation method of the input reading and initial antibody structure construction process is as follows: Step 1-1: Input data reading and parsing. Read the three-dimensional structure of the antigen epitope and obtain its residues and corresponding coordinate information; at the same time, read the known amino acid sequence and IMGT number of the antibody, and extract the three-dimensional coordinates of the known main chain atoms of the antibody, parse and organize them into the structured input required for subsequent modeling. Steps 1-2: Construction of the initial antibody structure representation, including the antibody... Represented as a graph structure ,in For a set of residue nodes, This is a set of edges connecting residues. Specifically, each antibody residue is treated as a node. initialize node features This feature is encoded by the residue type and its three-dimensional coordinates of all atoms. The basic information is used to construct the set; then, the edge set is constructed based on the spatial distance relationship between the residues. If the initial main chain coordinate distance between two residues does not exceed a preset threshold. If the two are connected, then an edge is formed. Finally, the initial structure of the antibody is obtained: .
3. The antibody design method based on flow matching and graph neural networks as described in claim 2, characterized in that, The specific implementation method of the antigen epitope surface modeling process is as follows: Step 2-1: Generate antigen surface point cloud. Generate surface point cloud data based on the three-dimensional atomic coordinates of the antigen molecule, and output a set of vertex coordinates located on the molecular surface. Step 2-2: Divide the surface point cloud by residue. Based on the correspondence between surface points and their corresponding atoms, assign the generated surface vertices to their corresponding antigen residues to obtain the surface vertex set for each residue: , where each surface point The three-dimensional spatial coordinates are represented as , This represents the actual number of surface vertices contained in the residue. Steps 2-3: Fixed-number vertex sampling and padding, performing a fixed number of samples for each residue: when At that time, from the set A fixed number of random samples One vertex; when When the time is right, all existing vertices are retained, and the remaining parts are filled with zero vectors; The final standardized set of surface vertices is obtained as follows: .
4. The antibody design method based on flow matching and graph neural networks as described in claim 3, characterized in that, The specific implementation of the local generation process of the binding region based on flow matching is as follows: Step 3-1: Construct the conditional input for flow matching generation. Use the 3D coordinates of the paratope region as the object for generating the flow matching ODE. Use its initial node features, coordinates, and the set of antigen surface vertices obtained in Step 2 as conditions for each binding region node. ,structure As its moment The initial state; Step 3-2: Generate the local structure of the paratope using flow matching, treating the paratope coordinates as continuous variables that evolve over time, and introduce ODE: in A learnable flow-matched velocity field.
5. The antibody design method based on flow matching and graph neural networks as described in claim 4, characterized in that, The specific implementation process of global structure propagation based on surface-enhanced message encoder is as follows: Step 4-1: Construct a surface enhancement message encoder. First, utilize the antigen surface vertex set generated in Step 2. Based on the spatial relationship between paratope residues and surface vertices, a fine geometric message oriented towards the paratope is constructed. The fine geometric message is based on continuous geometric difference and orientation-sensitive operator to realize the local structural transfer from the antigen surface to the paratope. message functions oriented towards three-dimensional coordinates The relationship between surface information and paratope coordinates is established using the following expression: in Represents coordinate difference. For learnable transformations; Step 4-2: Feature and coordinate update of the surface enhancement message encoder. Joint update of features and coordinates is performed in the antibody graph structure to propagate the local geometric influence of the antigen surface point cloud to the global structure. Specifically: for an edge in the antibody graph... First, the node features , The node edge messages are obtained by jointly encoding the surface-enhanced geometric messages between them. Subsequently, each node aggregates the edge messages of all neighboring nodes and inputs them into the feature update function to obtain new node features. ; A geometric update is constructed based on the relative coordinate difference between a node and its neighbors, and then added to the current node coordinates to obtain the new coordinates. Implement coordinate updates; Step 4-3: Integrate the surface augmentation message encoder with the velocity field network, directly incorporating the features and coordinate update mechanism of the surface augmentation message encoder into the velocity field network of the flow matching framework. This enables unified structure updates that combine local driving and global coordination.
6. The antibody design method based on flow matching and graph neural networks as described in claim 5, characterized in that, The specific implementation process of the joint convergence generation of the sequence and structure is as follows: Velocity Field Network The structure of the paratope is continuously updated based on the current node features and 3D coordinates. Simultaneously, by leveraging a surface-enhanced message encoder and a variant of the isomorphic graph neural network integrated into the network, local geometric changes are propagated across the entire antibody, achieving synchronous adjustment of the overall structure. Over time, the antibody's atomic coordinates gradually converge to a stable conformation, and the node features also converge to the corresponding amino acid distribution, thus generating a consistent 3D antibody structure. and binding region sequence This completes the integrated generation of sequences and structures.