Method and device for constructing antibody affinity prediction model
By constructing an antibody affinity prediction model of hierarchical coding module and prediction network module, the problem of insufficient time-consuming and accurate in the prior art is solved, and high-throughput and efficient antibody affinity prediction is achieved, which is suitable for antibody design and screening.
Patent Information
- Application Number
- CN202510452124.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-11
- Publication Date
- 2025-08-12
AI Technical Summary
The existing antibody affinity prediction methods rely on cumbersome wet experimental technology, are time-consuming and labor-intensive, and the algorithm based on energy function lacks accuracy. Machine learning-based methods lack experimental structural data and effective characterization, making it difficult to achieve high-throughput screening and accurately predict antibody affinity after mutation.
A antibody affinity prediction model is constructed, and a hierarchical coding module and prediction network module are designed by obtaining the data set for preprocessing. The structure, residue type and context characteristics of the antibody antigen complex are extracted using the message delivery network and self-attention mechanism to predict the changes in the affinity of the antibody after mutation, avoiding direct dependence on the mutant structure.
High accuracy and high efficiency antibody affinity prediction was achieved, with the Pearson correlation coefficient of 0.841 and the root mean square error of 1.195 kcal/mole, which is significantly better than existing tools and reduces the cost and time of high-throughput antibody modification screening.
Smart Images
Figure CN120472976A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of antibody-antigen affinity prediction, and relates to a method and device for constructing an antibody affinity prediction model. Background Art
[0002] Antibodies are multimeric proteins with molecular-specific recognition capabilities. The optimization process aims to screen for antibodies that precisely bind and possess drug-like properties. Antibody-based immunotherapies have demonstrated superior performance compared to traditional treatments for specific cancers, with fewer side effects and drug resistance. Consequently, antibody drugs have consistently topped the global best-selling drug list for many years and are the fastest-growing pharmaceutical segment.
[0003] Currently, the development of antibody drugs still relies on cumbersome wet lab techniques to screen for mutant antibodies with enhanced binding affinity. While effective, these methods are time-consuming and labor-intensive; therefore, affinity prediction methods that can rapidly screen for antibodies with superior affinity are urgently needed. In recent years, two computational approaches have emerged: energy-based algorithms and machine learning-based prediction models. Both methods aim to predict the affinity of a given antibody, accelerating the high-throughput screening process in actual antibody development and significantly reducing costs.
[0004] Energy-based algorithms are designed based on empirical considerations, but current research on macromolecular protein interactions at the microscopic level is incomplete, limiting the accuracy of these algorithms. The biggest challenge with deep learning affinity prediction lies in the lack of experimental structural data and the inability to effectively characterize and utilize this limited data. Summary of the Invention
[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for constructing an antibody affinity prediction model that can effectively improve prediction accuracy and achieve high-throughput prediction.
[0006] The technical problem that the present invention also aims to solve is to provide a method for predicting the affinity of antibodies after mutation.
[0007] The final technical problem to be solved by the present invention is to provide a device comprising the prediction model.
[0008] Technical solution: In order to solve the above technical problems, the present invention provides a method for constructing an antibody affinity prediction model, comprising the following steps:
[0009] 1) Obtaining a dataset for constructing a prediction model and preprocessing all the data in the dataset to obtain a training set, a validation set, and a test set; the data in the dataset includes the structure of the antibody-antigen complex before mutation, information after mutation, and the corresponding experimentally measured ΔΔG 真实值 ;
[0010] 2) Design a prediction model architecture, including a hierarchical encoding module and a prediction network module; the hierarchical encoding module is used to encode the input data to obtain the antibody-antigen complex encoding; the prediction network module predicts the difference in the encoding of the antibody-antigen complex before and after mutation, and outputs the predicted affinity change value;
[0011] 3) Using the training set and validation set obtained in step 1) to train the prediction model architecture designed in step 2);
[0012] 4) Using the test set obtained in step 1) to evaluate the prediction results obtained in step 3) to determine a stable antibody affinity prediction model.
[0013] The data set in step 1) is derived from the mutant antibody database AB-BIND, the mutant protein database SKEMPI1.0, and / or the SKEMPI2.0 database; the preprocessing is to first remove redundant and incomplete data, then process each remaining data sample, and after processing each sample, divide the entire data set;
[0014] Specific implementation methods include:
[0015] 1.1) All residues are classified based on their distance from the ligand (for antibody residues, the ligand is the antigen; for antigen residues, the ligand is the antibody) and whether the residues themselves are standard amino acids. The classification results include core region residues, common region residues, and inactive residues.
[0016] 1.2) Mask the classified residues: When the residue belongs to the core region, the mask value is 2, and it will directly form a node in the subsequent hierarchical encoding step; when the residue belongs to the common region, the mask value is 1, and in the subsequent hierarchical encoding step, it will be aggregated with the adjacent common region residues into a node, that is, coarse-grained processing; invalid residues that do not belong to standard amino acids have a mask value of 0 and will be ignored in the subsequent hierarchical encoding step.
[0017] 1.3) Select the data that belongs to the gold standard dataset in this field as the test set, and divide the remaining data into a training set and a validation set in a ratio of 8:2;
[0018] 1.4) Perform data augmentation on the training data; use hypothetical reverse mutation to 真实值 For data samples with values between [-2, 2], a reverse mutation sample from mutant to wild type is introduced, and the corresponding ΔΔG 真实值 Take the opposite of the original sample.
[0019] Wherein, the core region residues in step 1.1) are effective and the distance from the surface of the interacting protein is The common region residues are effective but are within The invalid residues are non-standard residues.
[0020] Among them, the hierarchical encoding module in step 2) includes a structural feature module for extracting and encoding input data in the first step, a residue type feature module for extracting and encoding input data in the second step, and a context feature module for extracting and fusing input data in the third step.
[0021] Among them, the specific implementation method of the first step of extracting and encoding the structural features of the input data is to extract structure-related features and represent the structure-related features as an undirected graph, and then aggregate the edge features into node features layer by layer through a message passing network to organize them into structure-related features; the structure-related features include initial node features and edge features, the initial node features include the dihedral angles of the residue nodes, and the edge features include spatial distance features, spatial rotation features and residue sequence distance features between residues.
[0022] Among them, the residue type feature module of the second step extracts and encodes the input data. It encodes the residue type features based on the structural features, uses the embedding layer to encode common amino acids, and then uses another message passing network to fuse the residue type of the corresponding node into the structural features of the previous step.
[0023] Among them, the third step of extracting and encoding the context feature module of the input data is implemented by using a bidirectional gated recurrent unit, and then the learned context information is weighted to the encoding of the previous step through the self-attention mechanism to obtain the final antibody-antigen complex encoding, namely the hierarchical encoding module.
[0024] Preferably, the specific implementation method of extracting the structural features of the input data is to extract the structural features and represent the structural features as an undirected graph, and the expression of the undirected graph is:
[0025]
[0026] in, It is an encoding of the antibody-antigen complex in the form of an undirected graph; is the node of the undirected graph, corresponding to the residue features in the antibody-antigen complex; ε is the edge of the undirected graph, corresponding to the interaction features between the residues; The expression is:
[0027]
[0028] v1 is the first node of the undirected graph; mis the mth node of the undirected graph; m is the total number of residues in the antibody-antigen complex; the initial node feature of each residue is expressed as:
[0029]
[0030] in, is the residue internal dihedral angle, The expression is:
[0031]
[0032] in, is the angle formed by rotating the bond between the amino nitrogen and the α carbon about the axis; ψ is the angle formed by rotating the bond between the α carbon and the carbonyl carbon about the axis; ω is the angle formed by rotating the bond between the carbonyl carbon and the next nitrogen atom of the residue about the axis; the expression for ε is:
[0033] ε={e ij |1≤i,j≤m,i≠j;
[0034] Among them, e ij represents the edge between residue i and residue j, and its expression is:
[0035] Among them, E pos It is a sequence distance feature, and the position encoding in the Transformer model is used to process the distance. pos The expression is:
[0036]
[0037] D pos is the residue spatial distance, D pos The expression is:
[0038] D pos =RBF(||x i，α -x j，α ||);
[0039] Among them, x i,α C represents residue i a coordinate; is the spatial orientation between residues, The expression is:
[0040]
[0041] q(·) is a quaternion; O i is the spatial rotation matrix of residue i; O j is the spatial rotation matrix of residue j; the spatial rotation matrix is based on C aCalculated in the local coordinate system centered on the atom; T represents the transposition of the matrix;
[0042] Preferably, the vector expression of the residue type feature is:
[0043] S r =Embedding(S)
[0044] Among them, S r is a residue type feature; said S represents the complete sequence of the antibody-antigen complex;
[0045] The context feature is extracted by a bidirectional gated recurrent unit, and the vector expression of the context feature is:
[0046] S c =GRU(S r );
[0047] S c is a context feature; S r is the residue type feature.
[0048] The specific steps of encoding the structural features, residue type features and context features include:
[0049] The structural feature graph is encoded using a message passing network with parameter θ1:
[0050]
[0051] in, The initial state of the corresponding graph encoding, is the initial state of the node, l is the current network layer; L is the total number of layers in the message passing network; Graph encoding for layer l+1 is the code of layer l and edge features ε are obtained by propagating in a message-passing-network (MPN) with parameter θ1 It is out Structural encoding after the network;
[0052] Encode the residue type features; use a message passing network with L layers and parameter θ2 to further integrate the residue type and internal conformation features into the node features; the first L / 2 layers only integrate the type information of the corresponding node, and the last L / 2 layers integrate the type information of its k nearest neighbors:
[0053]
[0054] The context feature is encoded in the following manner:
[0055]
[0056] Wherein, the prediction network module in step 2) includes a multilayer perceptron module (Multilayer Perceptron, MLP) and a mapping prediction module; the multilayer perceptron module is used to learn potential interaction patterns from the antibody-antigen complex encoding, and the multilayer perceptron module includes four fully connected layers and corresponding ReLU activation layers and Dropout layers; the mapping prediction module includes four fully connected layers, which are used to predict the difference in the encoding of the antibody-antigen complex before and after mutation output by the multilayer perceptron, and output the final ΔΔG 预测值 .
[0057] The specific training steps in step 3) include:
[0058] 3.1) hierarchically encoding the data in the training set preprocessed in step 1) using an encoding module to obtain the corresponding antibody-antigen complex encoding; and outputting a predicted affinity change value based on the difference in encoding of the antibody-antigen complex before and after mutation using a prediction network module;
[0059] 3.2) Calculate the loss function using the predicted values corresponding to the training set obtained in step 3.1) and the true values in the training set, and update the model parameters through backpropagation;
[0060] 3.3) hierarchically encoding the data in the validation set in step 1) using a hierarchical encoding module to obtain the corresponding antibody-antigen complex encoding; and using a prediction network module to output a predicted affinity change value based on the difference in encoding of the antibody-antigen complex before and after mutation;
[0061] 3.4) Calculate the root mean square error and Pearson correlation coefficient between the predicted values and the true values of the validation set obtained in step 3.3) and record these two indicators for subsequent observation of the change trend;
[0062] 3.5) Repeat steps 3.1) to 3.4) until the trend of the root mean square error and the Pearson correlation coefficient obtained in step 3.4) converges.
[0063] Wherein, the evaluation indicators in step 4) are root mean square error and Pearson correlation coefficient;
[0064] The Pearson correlation coefficient was calculated using the SciPy library in Python;
[0065] The expression of the root mean square error is:
[0066]
[0067] in:
[0068] N is the total number of samples; z represents the zth sample;
[0069] The loss function is used to optimize the antibody affinity prediction model during evaluation. The expression of the loss function is:
[0070]
[0071] Wherein: n is the total number of samples in the training set; z represents the zth sample; the Adams optimizer is used in the optimization, and a learning rate scheduler is used to reduce the learning rate when the training gradient descent slows down.
[0072] While providing the construction method as described above, the present invention also provides an antibody affinity prediction model abCAN (Antibody-Antigen Complex Attention Network) constructed based on the aforementioned construction method, which has more complete input selection; unlike most models based on antibody sequences, the present invention takes into account the important role of antigen-antibody and the structure of the antigen-antigen interaction surface, encompasses all structural and sequence information of antigen-antibody, and regards the antibody-antigen as a binding complex as a whole; and unlike other models that consider structure, the present invention does not need to input the structure of the mutant, but only needs to input the structural information of the antigen-antibody before mutation and how to perform mutation modification, so as to predict the affinity change of the antibody after mutation, which is more convenient to use, and coarse-grained distal residues in the preprocessing stage, weakening the contribution of residues farther away from the interaction surface without destroying the overall folding conformation. The antibody affinity prediction model abCAN constructed by the present invention is more in line with biological significance when extracting features. The three types of features, structure, amino acid residues and sequence context, almost cover all contribution features related to the interaction, and each type selects the most suitable deep learning method for feature encoding; using undirected graph To organize all the encoded features to form the final encoding structure.
[0073] The present invention also provides a method for predicting antibody affinity based on the model constructed by the method. The method for predicting antibody affinity includes inputting the structure of the antigen-antibody complex of the antibody to be predicted before mutation and mutation information into the model to obtain the affinity data of the antibody after mutation.
[0074] The present invention also provides a device comprising an antibody affinity prediction model, the device comprising: a hierarchical encoding module and a prediction network module; the hierarchical encoding module is used to encode data to obtain an antibody-antigen complex encoding; the prediction network module predicts based on the difference in encoding before and after mutation of the antibody-antigen complex to output a prediction result of the affinity change value.
[0075] Beneficial effects: Compared with the prior art, the present invention has the following advantages: The present invention provides a method for constructing an antibody affinity prediction model and a device for an antibody affinity prediction model. The core function of the present invention is to predict the change in antibody affinity after mutation, without the need for batch prediction or measurement of mutant structures, and can be directly applied to high-throughput screening of advantageous antibodies, or verification steps of antibody design. At the same time, the hierarchical coding method adopted by the present invention cleverly avoids the use of the structure after mutation, which is more in line with the situation in actual applications where only the original experimental structure can be obtained, solving the problem of scarce experimental structure data, and enabling large-scale prediction of the affinity of mutant antibodies. The present invention takes the original antibody-antigen complex to be optimized and the information of the transformation (such as mutation sites, related residues, etc.) as input, and outputs the change value ΔΔG of the affinity after the transformation. 预测值 Tests showed that the Pearson correlation coefficient between the prediction results of the model of the present invention and the experimental results was 0.841, the root mean square error was 1.195 kcal / mol, and the inference time on the validation set was 4.17 seconds, significantly higher than the 53.69 seconds of the DDG Predictor model. The accuracy of the prediction of the affinity change after mutation and the running time were significantly better than existing tools and models, significantly reducing the cost of high-throughput antibody modification screening. BRIEF DESCRIPTION OF THE DRAWINGS
[0076] Figure 1 This is the overall architecture diagram of the abCAN model;
[0077] Figure 2 is the prediction result of abCAN on the validation set;
[0078] Figure 3 For case analysis, the prediction results and inference speed of abCAN are compared with those of the other two methods for two representative samples;
[0079] Figure 4 Comparison results of abCAN and the other two methods by cross-validation with structure elimination one by one. DETAILED DESCRIPTION
[0080] The principle of the present invention is: drawing on experience from previous similar work, abCAN (Antibody-Antigen Complex Attention Network) was built, an antibody affinity prediction model using deep learning; for a known antibody-antigen complex, the affinity change value after any mutation modification is predicted without the need for additional prediction or measurement of the mutant structure. An embedding layer is used to encode the antibody-antigen residue type, and at the same time, a hierarchical encoding method is innovatively proposed to fuse structural and sequence information, organizing multi-level features in a biologically reasonable way; cleverly addressing the problem of lack of structural data, so that the model has higher accuracy and better biological interpretability. It involves building a deep learning model abCAN for mutant antibody affinity prediction, specifically using a message passing network in conjunction with a self-attention mechanism to perform hierarchical encoding of antibody-related data, and outputting the predicted affinity change value after transmission through a prediction network composed of a multi-layer perceptron, etc.
[0081] The technical solutions in the embodiments of the present invention are described clearly and completely below with reference to the accompanying drawings.
[0082] Example 1 Construction of abCAN Antibody Affinity Prediction Model
[0083] This example provides a method for constructing an antibody affinity prediction model. The workflow of the model is as follows: Figure 1 As shown:
[0084] 1) The prediction model constructed by the present invention accepts the pre-mutation complex structure and mutation information as input. It first identifies the residues on the antibody-antigen interaction surface and establishes a mask; then it coarse-grains the distal residues and aggregates multiple residues into one node.
[0085] 2) The encoder then enters hierarchical encoding. First, structural features are extracted and organized into structural features. The dihedral angle of each residue node is calculated and represented as the initial node feature using trigonometric functions. The Euclidean distance between residues is then calculated, and the nearest k neighbors are converted to RBF representations as the spatial distance features between residues. Furthermore, the spatial rotation feature is calculated by calculating the spatial rotation matrix of two residues and then represented as a quaternion. The sequence distance, represented by the positional encoding method in the Transformer, together with the spatial distance and spatial rotation features, constitutes the edge feature. Finally, through a message passing network with parameter θ1, the edge features are aggregated layer by layer into the node features and organized into structural features.
[0086] 3) The second step of hierarchical encoding is to encode residue features based on structural features. First, an embedding layer is used to encode 20 common amino acids, each represented by a learnable vector. Then, another message passing network with a parameter of θ2 is used to integrate the residue type of the corresponding node into the structural features from the previous step. The subsequent layers of the network integrate the residue types of neighboring nodes to learn interactions between residues.
[0087] 4) The final step of the hierarchical encoding is to incorporate sequence-level contextual information to capture long- and short-range interactions. This is achieved using a bidirectional gated recurrent unit (BGRU). The learned contextual information is then weighted into the encoding from the previous step via a self-attention mechanism. This step assigns higher attention to residues in interacting regions, resulting in the final complex encoding.
[0088] 5) The resulting complex code enters the prediction network. It is first propagated through a multilayer perceptron to learn the potential interaction patterns between the antibody and antigen. The complex codes before and after the mutation are propagated and then subtracted to produce a difference code, corresponding to the difference in interaction. Finally, the difference code enters the prediction network composed of fully connected layers, which outputs the predicted ΔΔG value.
[0089] The method for constructing the prediction model of the present invention specifically includes the following steps:
[0090] 1) Obtaining a dataset for constructing a prediction model and preprocessing all the data in the dataset to obtain a training set, a validation set, and a test set, and performing data enhancement on the training set; the data in the dataset includes the structure of the antibody-antigen complex before mutation, information after mutation, and the corresponding experimentally measured ΔΔG 真实值 ;
[0091] 2) Design a prediction model architecture, including a hierarchical encoding module and a prediction network module; the hierarchical encoding module is used to encode the input data to obtain the antibody-antigen complex encoding; the prediction network module predicts the difference in the encoding of the antibody-antigen complex before and after mutation, and outputs the predicted affinity change value;
[0092] 3) Using the training set and validation set obtained in step 1) to train the prediction model architecture designed in step 2);
[0093] 4) Using the test set obtained in step 1) to evaluate the prediction results obtained in step 3) to determine a stable antibody affinity prediction model.
[0094] Exemplarily, the source of the data set in step 1) is the mutant antibody database AB-BIND, the mutant protein database SKEMPI1.0, and the SKEMPI2.0 database; the preprocessing is to first remove redundant and incomplete data, then process each remaining data sample, and after processing each sample, divide the entire data set: specifically comprising the following steps:
[0095] 1.1) All residues are classified based on their distance from the ligand (for antibody residues, the ligand is the antigen; for antigen residues, the ligand is the antibody) and whether the residues themselves are standard amino acids. The classification results include core region residues, common region residues, and inactive residues.
[0096] 1.2) Mask the classified residues: When the residue belongs to the core region, the mask value is 2, and it will directly form a node in the subsequent hierarchical encoding step; when the residue belongs to the common region, the mask value is 1, and in the subsequent hierarchical encoding step, it will be aggregated with the adjacent common region residues into a node, that is, coarse-grained processing; invalid residues that do not belong to standard amino acids have a mask value of 0 and will be ignored in the subsequent hierarchical encoding step.
[0097] 1.3) Select the data that belongs to the gold standard dataset in this field as the test set, and divide the remaining data into a training set and a validation set in a ratio of 8:2;
[0098] 1.4) Perform data augmentation on the data in the training set; for example, use hypothetical reverse mutation to enhance ΔΔG 真实值 For data samples with values between [-2, 2], a reverse mutation sample from mutant to wild type is introduced, and the corresponding ΔΔG 真实值 Take the opposite of the original sample.
[0099] Exemplarily, the core region residues in step 1.1) are effective and are at a distance from the surface of the interacting protein. The common region residues are effective but are within The invalid residues are non-standard residues.
[0100] Among them, the hierarchical encoding modules in step 2) include a structural feature module for extracting and encoding input data in the first step, a residue type feature module for encoding in the second step, and a context feature module for fusing in the third step.
[0101] Among them, the specific implementation method of the first step of extracting and encoding the structural features of the input data is to extract structure-related features and represent the structure-related features as an undirected graph, and then aggregate the edge features into node features layer by layer through a message passing network to organize them into structure-related features; the structure-related features include initial node features and edge features, the initial node features include the dihedral angles of the residue nodes, and the edge features include spatial distance features, spatial rotation features and residue sequence distance features between residues.
[0102] Among them, the second step of encoding residue type features is to encode residue type features based on structural features. First, an embedding layer is used to encode common amino acids, and then another message passing network is used to fuse the residue type of the corresponding node into the structural features of the previous step.
[0103] Among them, the third step of encoding and fusing context features is achieved by using a bidirectional gated recurrent unit, and then the learned context information is weighted into the encoding of the previous step through a self-attention mechanism to obtain the final composite encoding, namely the hierarchical encoding module.
[0104] Exemplarily, a specific implementation of extracting the structural features of the input data is to extract structure-related features and represent the structure-related features as an undirected graph, where the expression of the undirected graph is:
[0105]
[0106] in, It is an encoding of the antibody-antigen complex in the form of an undirected graph; is the node of the undirected graph, corresponding to the residue features in the complex; ε is the edge of the undirected graph, corresponding to the interaction features between residues; The expression is:
[0107]
[0108] v1 is the first node of the undirected graph; m is the mth node of the undirected graph; m is the total number of residues in the complex; the initial node feature of each residue is expressed as:
[0109]
[0110] in, is the residue internal dihedral angle, The expression is:
[0111]
[0112] where φ is the angle formed by the bond between the amino nitrogen and the α-carbon; ψ is the angle formed by the bond between the α-carbon and the carbonyl carbon; ω is the angle formed by the bond between the carbonyl carbon and the next nitrogen atom in the residue; and ε is expressed as:
[0113] ε={e ij |1≤i,j≤m,i≠j;
[0114] Among them, e ij represents the edge between residue i and residue j, and its expression is:
[0115] Among them, E pos It is the sequence distance feature, and the position encoding in the Transformer model is used to process the sequence distance. pos The expression is:
[0116]
[0117] D pos is the residue spatial distance, D pos The expression is:
[0118] D pos =RBF(||x i，α -x j，α ||);
[0119] Among them, x i，α C represents residue i a coordinate; is the spatial orientation between residues, The expression is:
[0120]
[0121] q(·) is a quaternion; O i is the spatial rotation matrix of residue i; O j is the spatial rotation matrix of residue j; the spatial rotation matrix is based on C a Calculated in the local coordinate system centered on the atom; T represents the transposition of the matrix;
[0122] The vector expression of the residue type feature is:
[0123] S r =Embedding(S)
[0124] Among them, S r is a residue type feature; said S represents the complete sequence of the antibody-antigen complex;
[0125] The context feature is extracted by a bidirectional gated recurrent unit, and the vector expression of the context feature is:
[0126] S c =GRU(S r );
[0127] S c is a context feature; S r is the residue type feature.
[0128] The step of encoding the structural features and sequence features hierarchically includes:
[0129] The structural feature graph is encoded using a message passing network with parameter θ1:
[0130]
[0131] in, The initial state of the corresponding graph encoding, is the initial state of the node, l is the current network layer; L is the total number of layers in the message passing network; Graph encoding for layer l+1 is the code of layer l and edge features ε are obtained by propagating in a message-passing-network (MPN) with parameter θ1 It is out Structural encoding after the network;
[0132] Encode the residue type features; use a message passing network with L layers and parameter θ2 to further integrate the residue type and internal conformation features into the node features; the first L / 2 layers only integrate the type information of the corresponding node, and the last L / 2 layers integrate the type information of its k nearest neighbors:
[0133]
[0134] The context feature is encoded in the following manner:
[0135]
[0136] This invention innovatively proposes a hierarchical coding framework that integrates three types of features step by step, and uses a message passing network to fuse the three types of features of the complex step by step; at the same time, it uses an attention mechanism to capture the effects of residues connected by peptide bonds at the sequence level, giving higher attention weights to the interaction surfaces.
[0137] Wherein, the prediction network module in step 2) includes a multilayer perceptron module and a mapping prediction module; the multilayer perceptron module is used to learn the potential interaction pattern from the complex encoding, and the multilayer perceptron module includes four fully connected layers and corresponding ReLU activation layers and Dropout layers; the mapping prediction module includes four fully connected layers, which are used to predict the encoding difference of the complex before and after mutation output by the multilayer perceptron and output the final ΔΔG 预测值 .
[0138] The step of encoding the structural features and sequence features hierarchically includes:
[0139] The structural feature graph is encoded using a message passing network with parameter θ1:
[0140]
[0141] in, The initial state of the corresponding graph encoding, is the initial state of the node, l is the current network layer; L is the total number of layers in the message passing network; Graph encoding for layer l+1 is the code of layer l and edge features ε are obtained by propagating in a message-passing-network (MPN) with parameter θ1 It is out Structural encoding after the network;
[0142] Encode the residue type features; use a message passing network with L layers and parameter θ2 to further integrate the residue type and internal conformation features into the node features; the first L / 2 layers only integrate the type information of the corresponding node, and the last L / 2 layers integrate the type information of its k nearest neighbors:
[0143]
[0144] 3.3.3) Encoding the context features, wherein the encoding method is:
[0145]
[0146] It is an antibody-antigen complex encoding in the form of an undirected graph, attention is the attention mechanism, and GRU is the gated-recurrent-unit.
[0147] The specific training steps in step 3) include:
[0148] 3.1) hierarchically encoding the data in the training set preprocessed in step 1) using an encoding module to obtain the corresponding antibody-antigen complex encoding; and outputting a predicted affinity change value based on the difference in encoding of the antibody-antigen complex before and after mutation using a prediction network module;
[0149] 3.2) Calculate the loss function using the predicted values corresponding to the training set obtained in step 3.1) and the true values in the training set, and update the model parameters through backpropagation;
[0150] 3.3) hierarchically encoding the data in the validation set in step 1) using a hierarchical encoding module to obtain the corresponding antibody-antigen complex encoding; and using a prediction network module to output a predicted affinity change value based on the difference in encoding of the antibody-antigen complex before and after mutation;
[0151] 3.4) Calculate the root mean square error and Pearson correlation coefficient between the predicted values and the true values of the validation set obtained in step 3.3) and record these two indicators for subsequent observation of the change trend;
[0152] 3.5) Repeat steps 3.1) to 3.4) until the trend of the root mean square error and the Pearson correlation coefficient obtained in step 3.4) converges.
[0153] Wherein, the evaluation indicators in step 4) are root mean square error and Pearson correlation coefficient;
[0154] The Pearson correlation coefficient was calculated using the SciPy library in Python;
[0155] The expression of the root mean square error is:
[0156]
[0157] in:
[0158] N is the total number of samples; z represents the zth sample;
[0159] The loss function is used to optimize the antibody affinity prediction model during evaluation. The expression of the loss function is:
[0160]
[0161] Wherein: n is the total number of samples in the training set; z represents the zth sample; the Adams optimizer is used in the optimization, and a learning rate scheduler is used to reduce the learning rate when the training gradient descent slows down.
[0162] The ΔΔG value of the present invention reflects the difference in antibody-antigen binding energy ΔG. 预测值The predicted value of ΔΔG is output after the antibody-antigen complex is encoded by the prediction network module. In short, the prediction network model used in the present invention uses a multi-layer perceptron to learn the encoding of each antibody-antigen complex, calculate the difference encoding, and obtain the final prediction value through multiple fully connected layer mapping. 真实值 The ΔΔG values obtained from experiments in the database of mutant proteins used in the present invention are 真实值 .
[0163] Example 2
[0164] This embodiment provides a method for predicting antibody affinity based on the model constructed by the method described above. The method for predicting antibody affinity includes inputting the structure of the antibody-antigen complex before mutation and mutation information to be predicted into the model to obtain affinity data of the antibody after mutation.
[0165] Example 3
[0166] This embodiment provides a device including an antibody affinity prediction model, comprising: a hierarchical encoding module and a prediction network module; the hierarchical encoding module is used to encode data to obtain an antibody-antigen complex encoding; the prediction network module predicts the difference in the encoding of the antibody-antigen complex before and after mutation, thereby outputting a predicted result of the affinity change value.
[0167] Example 4
[0168] Figure 2 The prediction results of the abCAN prediction model constructed in Example 1 of the present invention on the validation set are shown. The horizontal axis is the true value and the vertical axis is the predicted value. It can be seen that the prediction regression results of the abCAN prediction model of the present invention are good, and most of the prediction points fall around the regression curve. Three other similar methods were compared and tested on this validation set. The results are shown in Table 1. The gold standard dataset comes from MutaBind2 (https: / / lilab.jysw.suda.edu.cn / research / mutabind2 / ), DDG predictor https: / / github.com / HeliXonProtein / binding-ddg- predictor and
[0169] GeoPPI(https: / / github.com / Liuxg16 / GeoPPI).
[0170] Table 1 Comparison of prediction results of different models on the gold standard dataset (test set)
[0171]
[0172] in: The results come from the web server; The results are from a local deployment of the source code.
[0173] Example 5 Ablation Experiment
[0174] To ensure that each feature contributes to the prediction, this embodiment performs an ablation experiment on the model constructed in Example 1. The results are shown in Table 2.
[0175] Table 2. Ablation results of different features or architectures of hierarchical encoding
[0176]
[0177] In ablation experiments, key features such as residue type are retained, while the absence of any other feature affects the results. Each feature encoding element contributes to performance, demonstrating that each selected feature makes a significant contribution to the interaction. Even if all features are retained, the hierarchical encoding network architecture lacks its core innovation. It simply stacks features without distinguishing between structural and sequence levels, and uses conventional network coding. The overall performance will still decline significantly, demonstrating that a good encoding architecture is crucial for integrating effective features.
[0178] Example 6 Verification of the Prediction Model of the Present Invention
[0179] The model constructed in Example 1 of the present invention was validated using the antibody-antigen complex 1EAW and the antibody-antigen complex 2B2X. Here, 1EAW and 2B2X refer to the protein complex numbering in the Protein Data Bank (PDB). The chain numbering (e.g., A chain, L chain, etc.) and residue positions (e.g., position 138, position 188) described below follow the PDB numbering format to facilitate user replication.
[0180] Figure 3 a shows the local conformational comparison of the interaction region before and after the single-point mutation of the antibody-antigen complex 1EAW. The horizontal axis of the bar graph represents three types of calculation methods or experimental data. The mutation site is at amino acid 138 of the A chain, which mutates from hydrophilic histidine with a large aromatic ring side chain to hydrophobic alanine with only a methyl side chain. The reduction of steric hindrance and the change in hydrophilicity result in an additional hydrogen bond inside glutamine 188 and threonine 139, thereby enhancing the overall affinity, ΔΔG 真实值 =-0.014<0. Figure 3b shows the local conformational comparison of the interaction region before and after the two-point mutation of the antibody-antigen complex 2B2X. The horizontal axis of the bar graph represents three types of calculation methods or experimental data; the mutation sites are at amino acids 28 and 52 of the L chain, one mutated from glutamine to serine, and the other mutated from tyrosine to asparagine. The mutation at position 28 did not lead to a change in the non-covalent interaction, while the mutation at position 52 destroyed the original π-π stacking interaction between the two benzene rings of glutamine at position 52 and tyrosine at position 48; but a hydrogen bond was generated. The overall affinity change can be expressed by ΔΔG 真实值 =0.682>0 indicates a decrease, meaning the lost π-π stacking interactions are stronger than the gained hydrogen bonds. The present invention successfully predicted a similar result, while the DDG predictor gave a large negative value, likely due to interference from newly formed hydrogen bonds. Hydrogen bonding is the primary non-covalent interaction and is crucial for protein interaction analysis. The present invention not only accounts for hydrogen bonding but also predicts the lost π-π stacking interactions, resulting in a more accurate analysis. As can be seen from the blue bar graph, the present invention's prediction speed is also significantly faster than GeoPPI.
[0181] To demonstrate the robustness of the model, a structure-by-structure elimination cross-validation was performed, excluding all mutation samples corresponding to a complex from the test set each time, and then calculating the Pearson correlation coefficient. Figure 4 The following table shows the comparison results, with the horizontal axis corresponding to the three methods. It can be seen that the prediction model abCAN of the present invention has a higher overall correlation coefficient and fewer outliers, indicating that no specific structure significantly affects model performance. While the other two methods require the predicted structure of the mutant as input, the prediction model abCAN can make more accurate predictions based solely on the original antibody-antigen complex structure. This demonstrates that the prediction model abCAN can make stable predictions based on less input information, unaffected by the specific complex structure.
Claims
1. A method for constructing an antibody affinity prediction model, characterized in that: The following steps are involved: 1) Obtaining a dataset for constructing a prediction model and preprocessing all the data in the dataset to obtain a training set, a validation set, and a test set; the data in the dataset includes the structure of the antibody-antigen complex before mutation, information after mutation, and the corresponding experimentally measured ΔΔG 真实值 ; 2) Design a prediction model architecture, including a hierarchical encoding module and a prediction network module; the hierarchical encoding module is used to encode the input data to obtain the antibody-antigen complex encoding; the prediction network module predicts the difference in the encoding of the antibody-antigen complex before and after mutation to output the predicted affinity change value ΔΔG 预测值 ; 3) Using the training set and validation set obtained in step 1) to train the prediction model architecture designed in step 2); 4) Using the test set obtained in step 1) to evaluate the prediction results obtained in step 3) to determine a stable antibody affinity prediction model.
2. The method for constructing an antibody affinity prediction model according to claim 1, wherein The source of the data set in step 1) is the mutant antibody database AB-BIND, the mutant protein database SKEMPI1.0 and / or SKEMPI2.0 database; the preprocessing is to first remove redundant and incomplete data, then process each remaining data sample, and after processing each sample, divide the entire data set.
3. The method for constructing an antibody affinity prediction model according to claim 1, wherein The hierarchical encoding modules in step 2) include a structural feature module for extracting and encoding input data in the first step, a residue type feature module for extracting and encoding input data in the second step, and a context feature module for extracting and encoding input data in the third step.
4. The method for constructing an antibody affinity prediction model according to claim 3, wherein: The specific implementation method of the structural feature module of the first step of extracting and encoding input data is to extract structure-related features and represent the structure-related features as an undirected graph, and then aggregate the edge features into node features layer by layer through a message passing network to organize them into structure-related features; the structure-related features include initial node features and edge features, the initial node features include the dihedral angles of residue nodes, and the edge features include spatial distance features, spatial rotation features and residue sequence distance features between residues.
5. The method for constructing an antibody affinity prediction model according to claim 3, wherein: The second step of extracting and encoding residue type feature modules encodes residue type features based on structural features, uses an embedding layer to encode common amino acids, and then uses another message passing network to fuse the residue type of the corresponding node into the structural features of the previous step.
6. The method for constructing an antibody affinity prediction model according to claim 3, wherein: The third step of extracting and encoding the context feature module of the input data is implemented by using a bidirectional gated recurrent unit, and then the learned context information is weighted to the encoding of the previous step through a self-attention mechanism to obtain the final antibody-antigen complex encoding before and after mutation, that is, the hierarchical encoding module.
7. The method for constructing an antibody affinity prediction model according to claim 1, wherein: The prediction network module in step 2) includes a multilayer perceptron module and a mapping prediction module; the multilayer perceptron module is used to learn potential interaction patterns from the differences in encoding before and after the antibody-antigen complex mutation, and the multilayer perceptron module includes four fully connected layers and corresponding ReLU activation layers and Dropout layers; the mapping prediction module includes four fully connected layers, which are used to predict the differences in encoding before and after the antibody-antigen complex mutation output by the multilayer perceptron and output the final ΔΔG 预测值 .
8. The method for constructing an antibody affinity prediction model according to claim 1, wherein The specific training steps in step 3) include: 3.1) hierarchically encoding the data in the training set preprocessed in step 1) using a hierarchical encoding module to obtain the corresponding antibody-antigen complex encoding; and using a prediction network module to output a predicted value of the affinity change value based on the difference in the encoding of the antibody-antigen complex before and after mutation; 3.2) Calculate the loss function using the predicted values corresponding to the training set obtained in step 3.1) and the true values in the training set, and update the model parameters through backpropagation; 3.3) hierarchically encoding the data in the validation set in step 1) using a hierarchical encoding module to obtain the corresponding antibody-antigen complex encoding; and using a prediction network module to output a predicted affinity change value based on the difference in encoding of the antibody-antigen complex before and after mutation; 3.4) Calculate the root mean square error and Pearson correlation coefficient between the predicted values and the true values of the validation set obtained in step 3.3) and record these two indicators for subsequent observation of the change trend; 3.5) Repeat steps 3.1) to 3.4) until the trend of the root mean square error and the Pearson correlation coefficient obtained in step 3.4) converges.
9. A method for predicting antibody affinity based on a model constructed by the method according to any one of claims 1 to 8, characterized in that: The method for predicting antibody affinity includes inputting the antigen-antibody complex structure of the antibody to be predicted before mutation and mutation information into a prediction model to obtain the change data of the antibody affinity after mutation.
10. A device comprising an antibody affinity prediction model, characterized in that: The device includes: a hierarchical encoding module and a prediction network module; the hierarchical encoding module is used to encode data to obtain an antibody-antigen complex encoding; the prediction network module predicts the difference in encoding before and after the antibody-antigen complex mutation to output a prediction result of the antibody affinity change value after the mutation.