A method and model for predicting antibody-antigen binding probability based on dilated convolution and channel attention mechanism and application thereof

Abstract

The application belongs to the field of biological medicine, and provides a method and model for predicting antibody-antigen binding possibility based on dilated convolution and channel attention mechanism, and application thereof. The method is based on dilated convolution and channel attention mechanism for predicting antibody-antigen binding possibility, a deep learning model is constructed by collecting a training data set, dilated convolution is introduced in training to expand the receptive field, and the channel attention mechanism is combined to enhance the feature expression capability, so as to realize efficient prediction. The application effectively expands the receptive field by dilated convolution to capture long program sequence dependence, combines the channel attention mechanism to adaptively strengthen key features, significantly improves the prediction accuracy under the condition of limited labeled data, and can be widely applied in the fields of antibody screening, drug design and immunotherapy.

Description

Technical Field

[0001] This invention belongs to the field of biomedicine, specifically relating to a method and model for predicting the probability of antibody-antigen binding based on dilated convolution and channel attention mechanisms, and its application. Background Technology

[0002] In the field of antibody development, traditional experimental methods have long been the main approach, including techniques for directed evolution such as immunization and phage or yeast display.

[0003] CN202210797823.6 discloses a method for panning a phage display library to obtain bispecific antibodies. Specifically, the phage display library panning method includes the following steps: extracting total RNA from human PBMCs, using the total RNA as a template to reverse transcribe and synthesize cDNA, using the cDNA as a template for multiple PCR amplifications, then digesting with enzymes to obtain linearized vector fragments and antibody gene fragments, ligating the fragments to obtain ligation products, introducing them into host cells for phage display to obtain a fully human single-domain antibody phage display antibody library, and then panning and performing a single-clone ELISA on the library.

[0004] CN201110206801.X discloses a method for analyzing the antigenic epitopes of monoclonal antibodies using a yeast surface display system and its application in vaccine development, comprising the following steps: (1) constructing a DNA library of the target protein; (2) displaying the DNA library on the surface of yeast to obtain a yeast display library; (3) mixing the monoclonal antibody of the target protein with yeast display library A and sorting out the yeast that binds to the monoclonal antibody; (4) extracting plasmids from the screened yeast and sequencing them to analyze the antigenic epitopes of the target protein.

[0005] While traditional antibody development techniques have achieved success in antibody screening, they also have several limitations, such as: 1. In ASC screening, due to significant individual differences in immune responses, the diversity of responses of B cells from different donors to the same antigen epitope is limited. This means that ASC clones can only capture a portion of potential binding sites, requiring the isolation of ASCs from a large number of subjects to cover a sufficient antibody spectrum, resulting in high costs and long cycles. 2. Uncontrollable epitope selection: The immune system tends to target strong immunogenic epitopes on the antigen surface, potentially ignoring conserved but crucial weak immunogenic epitopes, leading to screened antibodies failing to cover ideal targets. 3. Poor adaptability to multi-target scenarios: Some pathogens have hundreds of antigens, making it difficult to efficiently screen broad-spectrum neutralizing antibodies using traditional methods. 4. Time-consuming and low-throughput subsequent optimization: Screened antibodies may require further research and development to improve their potency, but structural information on antigen-antibody interactions relies on techniques such as cryo-electron microscopy, which have low epitope mapping throughput and cannot quickly guide antibody engineering. These methods suffer from high R&D investment, long screening cycles, and low throughput. Therefore, developing new rapid antibody screening methods holds significant promise for future applications. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method and model for predicting antibody-antigen binding probability based on dilated convolution and channel attention mechanisms, and its applications. This invention overcomes the problems of long cycles, high costs, and poor antibody binding specificity of traditional screening methods through artificial intelligence-assisted screening, enabling rapid and accurate screening of antibodies with significant binding ability to the target and blocking function.

[0007] To achieve this objective, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for predicting the probability of antibody-antigen binding based on dilated convolution and channel attention mechanisms, the method comprising:

[0009] (1) Data preprocessing steps: Obtain the heavy chain amino acid sequence of the antibody, the light chain amino acid sequence and the amino acid sequence of the antigen, encode the amino acid sequence and convert the amino acid sequence into a fixed-length one-hot feature matrix;

[0010] (2) Feature extraction step: The one-hot feature matrix is ​​input into the encoder, which is composed of several dilated convolution modules stacked sequentially; the matrix output after dilation convolution is enhanced by residual connection and random deactivation to prevent overfitting.

[0011] (3) Attention weighting and feature aggregation steps: Apply channel attention mechanism to the matrix processed by the feature extraction step to perform attention weighting to enhance feature expression, output channel weight vector, and multiply the weight vector with the original features channel by channel; perform global pooling on the feature maps output by the heavy chain, light chain and antigen coding branches after attention weighting to obtain fixed-length feature vectors, and then concatenate the three to form a fusion feature representation, forming a unified fusion feature matrix to provide input for subsequent classification prediction;

[0012] (4) Classification and prediction step: Classify and predict the fusion feature matrix and output the predicted value of the probability of antibody binding to antigen;

[0013] (5) Model training and optimization steps: Obtain positive and negative sample data as training set and validation set. Positive sample data refers to antibody sequences that can bind to antigens, and negative sample data refers to antibody sequences that cannot bind to antigens. Based on the above-labeled positive and negative sample data, use cross-entropy loss function and optimizer for training. Save model parameters when the validation set accuracy and loss rate meet preset conditions. The preset conditions are that when the validation accuracy is greater than or equal to the historical best value and the validation loss is less than or equal to the historical minimum value, the current model parameters are saved as the optimal model parameters.

[0014] Preferably, in the data preprocessing step, one-hot encoding is used to digitally represent the amino acid sequence, wherein each amino acid is converted into a one-hot vector of length 20, corresponding to 20 standard amino acids; to ensure consistent input dimensions, sequences that are not long enough are padded to a uniform length; after encoding and padding, a single sequence is represented as a two-dimensional feature matrix composed of feature vectors of multiple amino acids; multiple encoded and padded sequences can be combined in batches to form model input features for subsequent feature extraction and learning.

[0015] Preferably, in the feature extraction step, the dilated convolution module includes several parallel convolutional branches with different dilation rates, used to extract local and global features of the sequence within different receptive fields.

[0016] Preferably, in the feature extraction step, the number of channels in each group of dilated convolution modules increases layer by layer to enhance the feature representation capability.

[0017] Preferably, in the feature extraction step, the output of each dilated convolution module employs residual connections and random deactivation to enhance feature transfer and prevent overfitting.

[0018] Through the aforementioned dilated convolution structure, the model can effectively capture the multi-scale features of antibody and antigen sequences with a limited number of layers, providing high-quality representations for subsequent channel attention weighting and classification prediction.

[0019] Preferably, the channel attention mechanism weighting and feature aggregation steps include: performing global pooling on the features of each channel in the matrix processed by the feature extraction step to obtain a channel description vector; then mapping it through two fully connected layers to output a channel weight vector; and multiplying the weight vector with the original features channel by channel to highlight key features, suppress redundant features, and improve feature selectivity.

[0020] Preferably, in the two-layer fully connected network, the first layer is activated using a linear rectified function (ReLU), and the second layer is activated using a sigmoid function.

[0021] Preferably, in the classification prediction step, the fused features in the fused feature matrix are input to the classification prediction module for classification processing.

[0022] Preferably, the classification prediction module is implemented through a fully connected classification module, which includes multi-layer feature mapping units and sets residual connection structures between each layer to enhance feature transfer and model stability.

[0023] Preferably, the feature mapping unit is equipped with a nonlinear activation and regularization mechanism to improve the nonlinear expression capability of the model and reduce the risk of overfitting; finally, a binding probability prediction unit is set at the model output end to output the binding probability between the antibody and the antigen.

[0024] Preferably, in the model training and optimization steps, binary crossentropy is used as the loss function, and the difference between the predicted probability value and the true label is used as the optimization objective.

[0025] During the optimization process, an optimization algorithm with an adaptive learning rate adjustment mechanism can be used to update parameters. The learning rate range and batch size can be adjusted according to the dataset size and model complexity.

[0026] In one specific embodiment of the present invention, the optimization algorithm is Adaptive Moment Estimation (Adam), the learning rate is set to 1e-3 to 1e-4, and the batch size is 12.

[0027] In this invention, the model selection and saving strategy involves real-time monitoring of the accuracy and loss value of the validation set during training. When the validation accuracy is greater than or equal to the historical best value and the validation loss is less than or equal to the historical minimum value, the current model parameters are saved as the optimal model to ensure that the model achieves the best balance between performance and generalization ability.

[0028] Secondly, the present invention provides a predictive model for antibody-antigen binding probability based on dilated convolution and channel attention mechanisms, the model comprising:

[0029] Data preprocessing module: acquires the heavy chain amino acid sequence of the antibody, the light chain amino acid sequence of the antigen, encodes the amino acid sequence, and converts the amino acid sequence into a fixed-length one-hot feature matrix;

[0030] Feature extraction module: The one-hot feature matrix is ​​input into the encoder, which is composed of several dilated convolution modules stacked sequentially; the matrix output after dilation convolution uses residual connection and random deactivation to enhance feature transfer and prevent overfitting.

[0031] Attention weighting and feature aggregation module: Apply channel attention mechanism to the matrix processed by the feature extraction step to perform attention weighting to enhance feature representation, output channel weight vector, and multiply the weight vector with the original features channel by channel; perform global pooling on the feature maps output by the three coding branches of the heavy chain, light chain and antigen after attention weighting to obtain fixed-length feature vectors, and then concatenate the three to form a fused feature representation, forming a unified fused feature matrix to provide input for subsequent classification prediction;

[0032] Classification and prediction module: performs classification and prediction on the fused feature matrix and outputs a predicted value of the probability of antibody-antigen binding;

[0033] Training and Optimization Module: Acquires positive and negative sample data as training and validation sets, respectively. Positive sample data refers to antibody sequences that can bind to antigens, while negative sample data refers to antibody sequences that cannot bind to antigens. Based on the labeled positive and negative sample data, training is performed using the cross-entropy loss function and an optimizer. The model parameters are saved when the validation accuracy and loss rate meet preset conditions. The preset conditions are that when the validation accuracy is greater than or equal to the historical best value and the validation loss is less than or equal to the historical minimum value, the current model parameters are saved as the optimal model parameters.

[0034] Preferably, in the data preprocessing module, one-hot encoding is used to digitally represent the amino acid sequence, where each amino acid is converted into a one-hot vector of length 20, corresponding to 20 standard amino acids; to ensure consistent input dimensions, sequences that are not long enough are padded to a uniform length; after encoding and padding, a single sequence is represented as a two-dimensional feature matrix composed of feature vectors of multiple amino acids; multiple encoded and padded sequences can be combined in batches to form model input features for subsequent feature extraction and learning.

[0035] Preferably, in the feature extraction module, the dilated convolution module includes several parallel convolutional branches with different dilation rates, used to extract local and global features of the sequence within different receptive fields.

[0036] Preferably, in the feature extraction module, the number of channels in each group of dilated convolution modules increases layer by layer to enhance the feature representation capability.

[0037] Preferably, in the feature extraction module, the output of each dilated convolution module adopts residual connection and random deactivation to enhance feature transfer and prevent overfitting.

[0038] Preferably, the channel attention mechanism weighting and feature aggregation steps include: performing global pooling on the features of each channel in the matrix processed by the feature extraction step to obtain a channel description vector; then mapping it through two fully connected networks to output a channel weight vector; and multiplying the weight vector with the original features channel by channel to highlight key features, suppress redundant features, and improve feature selectivity.

[0039] Preferably, the mapping of the two fully connected network includes activating the first layer using a linear rectified function (ReLU) and activating the second layer using a sigmoid function.

[0040] Preferably, in the classification prediction module, the fused features in the fused feature matrix are input to the classification prediction module for classification processing.

[0041] Preferably, the classification prediction module is implemented through a fully connected classification module, which includes multi-layer feature mapping units and sets residual connection structures between each layer to enhance feature transfer and model stability.

[0042] Preferably, the feature mapping unit is equipped with a nonlinear activation and regularization mechanism to improve the nonlinear expression capability of the model and reduce the risk of overfitting; finally, a binding probability prediction unit is set at the model output end to output the binding probability between the antibody and the antigen.

[0043] Preferably, in the model training and optimization module, binary crossentropy is used as the loss function, and the difference between the predicted probability value and the true label is used as the optimization objective.

[0044] Thirdly, the present invention provides the application of the method for predicting the probability of antibody-antigen binding based on dilated convolution and channel attention mechanism as described in the first aspect, or the prediction model for the probability of antibody-antigen binding based on dilated convolution and channel attention mechanism as described in the second aspect, in antibody screening.

[0045] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the method for predicting and screening antibodies based on dilated convolution and channel attention mechanisms for antibody-antigen binding probability as described in the first aspect.

[0046] Fifthly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and the processor executes the computer program to implement the steps of the method for predicting the probability of antibody-antigen binding based on dilated convolution and channel attention mechanism as described in the first aspect.

[0047] In a sixth aspect, the present invention provides an antibody that specifically binds to the TIGIT protein, the antibody comprising a heavy chain variable region and a light chain variable region;

[0048] The heavy chain variable region CDR1 of the monoclonal antibody is shown in SEQ ID NO:1, CDR2 in SEQ ID NO:2, and CDR3 in SEQ ID NO:3; the light chain variable region CDR1 of the monoclonal antibody is shown in SEQ ID NO:4, CDR2 in SEQ ID NO:5, and CDR3 in SEQ ID NO:6.

[0049] The antibody provided by this invention can specifically recognize and bind to the TIGIT protein and effectively block the interaction between TIGIT and its various ligands (including CD155 and Nectin-4). Compared with existing antibodies that can only block CD155, the antibody of this invention has a broader blocking spectrum and can simultaneously inhibit multiple immunosuppressive pathways, thereby significantly enhancing the immune activity of T cells and / or NK cells. Due to its high specificity and excellent functional blocking ability, the antibody of this invention has outstanding application advantages in tumor immunotherapy and the diagnosis and treatment of related diseases.

[0050] Preferably, the amino acid sequence of the heavy chain variable region of the antibody is shown in SEQ ID NO:7, and the amino acid sequence of the light chain variable region of the antibody is shown in SEQ ID NO:8.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] This invention proposes a method for predicting antibody-antigen binding probability based on dilated convolution and channel attention mechanisms. This method can significantly improve prediction accuracy during antibody screening, reduce antibody development and production costs, and greatly shorten the antibody discovery cycle, thereby providing an efficient and economical new technology path for the development of immune checkpoint antibody drugs. Attached Figure Description

[0053] Figure 1 This is a flowchart of a method for predicting the probability of antibody-antigen binding based on dilated convolution and channel attention mechanisms.

[0054] Figure 2This is a graph showing the protein binding capacity test results for sequences TIGIT-01 to TIGIT-09.

[0055] Figure 3 This is a graph showing the protein binding capacity test results for sequences TIGIT-10 to TIGIT-18.

[0056] Figure 4 This is a graph showing the protein binding capacity test results for sequences TIGIT-19 and TIGIT-20.

[0057] Figure 5 This is a graph showing the cell binding capacity test results for sequences TIGIT-01 to TIGIT-10.

[0058] Figure 6 This is a graph showing the cell binding ability test results for sequences TIGIT-11 to TIGIT-19.

[0059] Figure 7 This is a graph showing the results of the TIGIT-20 sequence cell binding ability assay.

[0060] Figure 8 This is a graph showing the blocking detection results of TIGIT-13 and its ligand CD155.

[0061] Figure 9 This is a graph showing the blocking detection results of TIGIT-13 and its ligand Nectin-4. Detailed Implementation

[0062] The technical solution of the present invention will be further illustrated below through specific embodiments. Those skilled in the art should understand that the embodiments described are merely illustrative of the present invention and should not be construed as limiting the invention in any way.

[0063] Where specific techniques or conditions are not specified in the examples, they shall be performed in accordance with the techniques or conditions described in the literature in this field, or in accordance with the product instructions. Reagents or instruments whose manufacturers are not specified are all conventional products that can be purchased through legitimate channels.

[0064] Example 1

[0065] This embodiment provides a method for predicting the probability of antibody-antigen binding based on dilated convolution and channel attention mechanisms.

[0066] This embodiment takes TIGIT antibody screening as an example, providing the protein binding and cell binding experiment results of 20 TIGIT antibodies and the screening process.

[0067] In this embodiment, an antibody-antigen association database was first constructed, containing various antigens and their experimentally validated binding and non-binding antibody sequences, to provide a source of positive and negative sample data for model training. Furthermore, an antibody library targeting the TIGIT antigen was constructed using Single B-cell sequencing technology. This antibody library underwent enrichment processing and contains a large number of diverse antibody sequences.

[0068] Based on the aforementioned database and antibody library, the deep learning prediction method proposed in this invention, based on dilated convolutional neural networks and channel attention mechanism, is used to perform batch screening of candidate antibody sequences. The method flow is as follows: Figure 1 Using this method, several candidate antibody sequences with a high probability of binding to the target antigen were obtained from the antibody library. The above results demonstrate that the method of the present invention can effectively predict the binding probability of antibodies to antigens under limited training data conditions, thus providing strong technical support for the development of novel antibody drugs.

[0069] 1. The deep learning model is a prediction model for the probability of antibody-antigen binding based on dilated convolution and channel attention mechanisms. The main architecture of the model includes:

[0070] (1) Data preprocessing module: Obtain the heavy chain amino acid sequence of the antibody, the light chain amino acid sequence and the amino acid sequence of the antigen, encode the amino acid sequence and convert the amino acid sequence into a fixed-length one-hot feature matrix.

[0071] Specifically, in the data preprocessing step, one-hot encoding is used to digitally represent the amino acid sequence, where each amino acid is converted into a one-hot vector of length 20, corresponding to 20 standard amino acids; to ensure consistent input dimensions, sequences that are not long enough are padded to a uniform length; after encoding and padding, a single sequence is represented as a two-dimensional feature matrix composed of feature vectors of multiple amino acids; multiple encoded and padded sequences can be combined in batches to form model input features for subsequent feature extraction and learning.

[0072] (2) Encoder module (i.e. feature extraction module): The one-hot feature matrix is ​​input to the encoder, which is composed of several dilated convolution modules stacked sequentially; the matrix output after dilation convolution uses residual connection and random deactivation to enhance feature transfer and prevent overfitting.

[0073] In the feature extraction step, each encoder consists of several dilated convolutional modules stacked sequentially. Each group of dilated convolutional modules contains multiple parallel convolutional branches, each using a different dilation rate to extract local and global features from different receptive fields. Preferably, the dilation rate of the convolutional branches can be set to incrementally increasing integers, such as 3, 9, and 12, to achieve multi-scale feature fusion. The number of channels in each group of convolutional modules increases layer by layer, for example, from 32 to 48, to enhance the model's expressive power. The kernel size is preferably 3 to achieve a balance between computational efficiency and feature resolution. After the output of each group of modules, residual connections and dropout are applied to enhance feature transfer and prevent overfitting. Through the above dilated convolutional structure, the model can effectively capture multi-scale features of antibody and antigen sequences with a limited number of layers, providing high-quality representations for subsequent channel attention weighting and classification prediction.

[0074] (3) Attention weighting and feature aggregation module: Apply channel attention mechanism to the matrix processed by the feature extraction step to perform attention weighting to enhance feature expression, output channel weight vector, multiply the weight vector with the original feature channel by channel; perform global pooling on the multiplied channel weight vector to obtain a fixed-length vector and concatenate it to form a unified fusion feature matrix.

[0075] In the attention weighting step, the channel attention mechanism can employ a Squeeze-and-Excitation (SE) structure. This structure generates weight vectors characterizing channel importance by performing global pooling and non-linear mapping operations on the features of each channel. Specifically, the features of each channel in the matrix processed by the feature extraction step are globally averaged to obtain a channel description vector. This vector is then passed through two fully connected layers; for example, the first layer uses a Rectified Linear Function (ReLU) activation, and the second layer uses a Sigmoid function activation, outputting a channel weight vector. This weight vector is then multiplied channel-by-channel with the original features, thereby highlighting key features, suppressing redundant features, and improving feature selectivity.

[0076] Specifically, the feature aggregation module includes performing global pooling (such as global max pooling or global average pooling) on ​​the feature maps output from the attention-weighted heavy chain, light chain, and antigen coding branches to obtain fixed-length feature vectors. These three vectors are then concatenated to form a fused feature representation, providing input for subsequent classification prediction.

[0077] (4) Classification prediction module: Classifies and predicts the fusion feature matrix and outputs the predicted value of the probability of antibody binding to antigen.

[0078] The classification prediction module is implemented through a fully connected classification module, which includes multi-layer feature mapping units and residual connection structures between each layer to enhance feature transfer and model stability. Each feature mapping unit incorporates non-linear activation and regularization mechanisms, such as sigmoid activation and dropout, to improve the model's non-linear expressive power and reduce overfitting risk. A binding probability prediction unit is included at the model output to output the binding probability between the antibody and the antigen.

[0079] (5) Training and optimization module: Obtain positive and negative sample data as training set and validation set. Positive sample data refers to antibody sequences that can bind to antigens, and negative sample data refers to antibody sequences that cannot bind to antigens. Based on the above-labeled positive and negative sample data, the cross-entropy loss function and optimizer are used for training. When the validation set accuracy and loss rate meet the preset conditions, the model parameters are saved. The preset conditions are that when the validation accuracy is greater than or equal to the historical best value and the validation loss is less than or equal to the historical minimum value, the current model parameters are saved as the optimal model parameters.

[0080] Specifically, in the model training and optimization steps, binary crossentropy is used as the loss function, and the difference between the predicted probability value and the true label is used as the optimization objective.

[0081] During the optimization process, the parameter settings include: the optimizer is Adam (adaptive moment estimation optimization algorithm), the learning rate range is 1e-3 to 1e-4, and the batch size is 12.

[0082] In the model selection and saving strategy, the accuracy and loss value of the validation set are monitored in real time during training. When the validation accuracy is greater than or equal to the historical best value and the validation loss is less than or equal to the historical minimum value, the current model parameters are saved as the optimal model parameters to ensure that the model achieves the best balance between performance and generalization ability.

[0083] 2. Training process.

[0084] In this embodiment, 10 known antibody sequences capable of binding to the target antigen TIGIT via protein ELISA were first selected as part of the positive samples. Simultaneously, other unrelated target antibodies, along with known protein-bound or non-bound antibody sequences, were further introduced, totaling 689 sequences, to enrich the training set and serve as negative or non-specific samples. Thus, the entire training data contained a total of 190 positive sequences and 499 negative sequences.

[0085] During training, this embodiment employs a prediction model based on dilated convolution and channel attention mechanisms to learn from the aforementioned data, continuously monitoring changes in the accuracy and loss rate of the validation set. Within a preset number of training epochs, when the validation accuracy is greater than or equal to the historical best value and the validation loss is less than or equal to the historical minimum value, the current model parameters are saved as the optimal model parameters. Ultimately, the obtained optimal model parameters are used to predict the probability of candidate antibody binding to antigen.

[0086] 3. Prediction process.

[0087] The best-preserved model was used to predict the binding probability of the single-B sequencing database. Based on the prediction results, 20 sequences with high predicted binding probabilities and those with average probabilities were selected for protein ELISA and cell ELISA binding assays, respectively. When selecting test sequences, the selected sequences differed from known binding sequences by 4 or more amino acids in the heavy chain CDR region, and the heavy chain CDR regions of the test sequences also differed from each other by 4 or more amino acids.

[0088] 4. Model prediction results and experimental verification results (Table 1).

[0089] Table 1

[0090]

[0091] In Table 1, "NA" indicates that the current column value of the sequence has not been measured. For example, the cell ELISA value will only be measured when the ELISA result of the protein to be tested is binding; otherwise, the cell ELISA value will not be measured when the ELISA result of the protein to be tested is not binding or weakly binding. In this case, the cell ELISA value corresponding to the sequence is recorded as NA.

[0092] The deep learning model described in this invention outputs a prediction result between 0 and 1, representing the probability of antibody binding to antigen. Twenty candidate antibody sequences were selected from the model's prediction results, with predicted binding probabilities ranging from 0.54 to 0.99, and these were experimentally verified.

[0093] 5. Protein ELISA experimental procedure.

[0094] ELISA binding assays were performed. Recombinant HumanTIGIT / his Protein (Cat#: BI120) carrying the his tag was used to analyze the binding with the expressed antibody. The antigen protein was coated overnight at 4°C with a concentration of 2 µg / mL in 96-well microplates, followed by blocking with 1% BSA for 2 hours at room temperature the next day. Then, serially diluted antibodies were added and incubated at room temperature for 1 hour. Next, anti-human IgG (Fc specific) (Cat#A0170-1mL, 1:5000) was added and incubated for 30 minutes at room temperature. Finally, TMB substrate was added, and the OD450 value was measured using a microplate reader. Results are as follows: Figure 2 , Figure 3 , Figure 4 As shown in the figure. The results indicate that 19 out of the 20 antibody sequences sampled exhibit protein-level binding with TIGIT, demonstrating the model's ability to screen for binding antibodies.

[0095] 6. Cell ELISA experimental procedure.

[0096] Cell binding was detected using flow cytometry (FACS). First, cultured cells were prepared to a concentration of 4 × 10⁶ cells / mL using MACS buffer. 6 Cells / ml were prepared by adding 50 μL of cell suspension to each well of a 96-well plate. The prepared antibody was serially diluted 4-fold with MACS buffer, and 50 μL of antibody dilution was added to each well. The plate was incubated at 4°C in the dark for 1 hour, followed by washing three times with MACS buffer. Cells were then resuspended in anti-human fluorescent secondary antibody (Jackson, Goat anti-human IgG AF647, Cat#109-605-190), incubated at 4°C in the dark for 30 minutes, washed three times with MACS buffer, and finally resuspended in 200 μL of MACS buffer. Flow cytometry was used for data analysis. Results are as follows: Figure 5 , Figure 6 , Figure 7 As shown in the figure. The results indicate that most of the 19 antibodies that exhibit protein-level binding activity also have cell-binding ability, with TIGIT-08, TIGIT-12, and TIGIT-13 showing strong cell-level binding ability to TIGIT.

[0097] As shown in Table 1 and Figure 2 , Figure 3 , Figure 4As shown, the binding assay results based on protein ELISA revealed that 19 out of the 20 candidate antibodies tested were able to bind to the TIGIT antigen, achieving a verification accuracy of 95%. This result demonstrates that the deep learning model described in this invention can effectively learn the binding characteristics of antibodies and antigens, and possesses practical value in assisting in the screening of antibodies with binding capabilities.

[0098] In summary, the method of this invention can significantly improve the efficiency and accuracy of antibody screening under limited experimental verification conditions, providing strong technical support for the development of TIGIT antibodies.

[0099] Example 2

[0100] This embodiment verifies the TIGIT-13 sequence screened in Example 1 through ligand blocking experiments.

[0101] The CDR region sequence of antibody TIGIT-13 is shown below:

[0102] Heavy chain CDR1: SEQ ID NO:1.

[0103] Heavy chain CDR2: SEQ ID NO:2.

[0104] Heavy chain CDR3: SEQ ID NO:3.

[0105] Light chain CDR1: SEQ ID NO:4.

[0106] Light chain CDR2: SEQ ID NO:5.

[0107] Light chain CDR3: SEQ ID NO:6.

[0108] Antibody TIGIT-13: Heavy chain variable region sequence: SEQ ID NO:7.

[0109] Antibody TIGIT-13: Light chain variable region sequence: SEQ ID NO:8.

[0110] 1. Ligand blocking experimental procedure.

[0111] Cell binding was detected using flow cytometry (FACS). First, cultured cells were prepared to a concentration of 4 × 10⁶ cells / mL using MACS buffer. 6Cells / ml were measured, and 50 μL of cell suspension was added to each well of a 96-well plate. Biotin-labeled TIGIT ligand protein (Biotin-Nectin-4hFc or Biotin-CD155hFc) was diluted to 200 nM with MACS buffer. The prepared antibody was serially diluted 4-fold with MACS buffer, and 25 μL of biotin-Nectin-4hFc protein and 25 μL of diluted antibody were added to designated wells respectively. The plate was incubated at 4°C in the dark for 1 hour, and washed three times with MACS buffer. Subsequently, the cells were resuspended with fluorescent secondary antibody (BD Pharmingen™ PE Streptavidin (1:1000), Cat#554061), incubated at 4°C in the dark for 30 minutes, washed three times with MACS buffer, and finally resuspended with 200 μL of MACS buffer. Flow cytometry was used for data analysis.

[0112] The ligand blocking assay was performed only on TIGIT-13, which showed the best performance in the cell binding assay. The results are as follows: Figure 8 , Figure 9 As shown. Figure 8 This indicates that the TIGIT-13 sequence has the ability to block the ligand CD155; Figure 9 The results showed that the TIGIT-13 sequence has the ability to block the ligand Nectin-4. Therefore, the TIGIT-13 antibody can not only bind to the TIGIT antigen, but also block the interaction between TIGIT and its ligands CD155 and Nectin-4, demonstrating potential functional activity and further drug development value.

[0113] Example 3

[0114] This embodiment provides a method for screening antibodies based on dilated convolution and channel attention mechanisms to predict the probability of antibody-antigen binding. The method includes the following steps:

[0115] 1. Data preparation and preprocessing: Obtain the amino acid sequences of the antibody heavy chain, light chain, and antigen, encode the amino acid sequences, and convert the amino acid sequences into a fixed-length one-hot feature matrix.

[0116] 2. Feature extraction: The one-hot feature matrix is ​​input into the encoder, which is composed of several dilated convolution modules stacked sequentially; the matrix output after dilation convolution uses residual connection and random deactivation to enhance feature transfer and prevent overfitting.

[0117] 3. Attention weighting and feature aggregation steps: Apply channel attention mechanism to the matrix processed in the feature extraction step to enhance feature representation, output channel weight vector, and multiply the weight vector with the original features channel by channel; perform global pooling on the multiplied channel weight vector to obtain a fixed-length vector and concatenate them to form a unified fusion feature matrix.

[0118] 4. Classification and prediction step: Perform classification and prediction on the fusion feature matrix, and output the predicted value of the probability of antibody-antigen binding.

[0119] 5. Model Training and Optimization Steps: Obtain positive and negative sample data as training and validation sets, respectively. Positive sample data refers to antibody sequences that can bind to antigens, and negative sample data refers to antibody sequences that cannot bind to antigens. Based on the above-labeled positive and negative sample data, train the model using the cross-entropy loss function and optimizer. Save the model parameters when the validation accuracy and loss rate meet preset conditions. The preset conditions are that when the validation accuracy is greater than or equal to the historical best value and the validation loss is less than or equal to the historical minimum value, the current model parameters are saved as the optimal model parameters.

[0120] Through the above steps, this embodiment can effectively predict the probability of antibody-antigen binding even with a limited number of samples, providing auxiliary support for antibody screening and related drug development.

[0121] The applicant declares that the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention fall within the protection and disclosure scope of the present invention.

Claims

1. A method for predicting the probability of antibody-antigen binding based on dilated convolution and channel attention mechanisms, characterized in that, The method includes: (1) Data preprocessing steps: Obtain the heavy chain amino acid sequence of the antibody, the light chain amino acid sequence and the amino acid sequence of the antigen, encode the amino acid sequence and convert the amino acid sequence into a fixed-length one-hot feature matrix; (2) Feature extraction step: The one-hot feature matrix is ​​input into the encoder, which is composed of several dilated convolution modules stacked sequentially; the matrix output after dilation convolution is enhanced by residual connection and random deactivation to prevent overfitting. (3) Attention weighting and feature aggregation steps: Apply channel attention mechanism to the matrix processed by the feature extraction step to perform attention weighting to enhance feature expression, output channel weight vector, and multiply the weight vector with the original features channel by channel; perform global pooling on the feature maps output by the heavy chain, light chain and antigen coding branches after attention weighting to obtain fixed-length feature vectors, and then concatenate the three to form a fusion feature representation, forming a unified fusion feature matrix to provide input for subsequent classification prediction; (4) Classification and prediction step: Classify and predict the fusion feature matrix and output the predicted value of the probability of antibody binding to antigen; (5) Model training and optimization steps: Obtain positive and negative sample data as training set and validation set. Positive sample data refers to antibody sequences that can bind to antigens, and negative sample data refers to antibody sequences that cannot bind to antigens. Based on the labeled positive and negative sample data, use cross-entropy loss function and optimizer for training. Save model parameters when the validation set accuracy and loss rate meet preset conditions. The preset conditions are that when the validation accuracy is greater than or equal to the historical best value and the validation loss is less than or equal to the historical minimum value, the current model parameters are saved as the optimal model parameters. In the data preprocessing step, one-hot encoding is used to digitally represent the amino acid sequence, where each amino acid is converted into a one-hot vector of length 20, corresponding to 20 standard amino acids. To ensure consistent input dimensions, sequences that are too short are padded to a uniform length. After encoding and padding, a single sequence is represented as a two-dimensional feature matrix composed of feature vectors of multiple amino acids. Multiple encoded and padded sequences can be combined in batches to form the model input features for subsequent feature extraction and learning.

2. The method for predicting antibody-antigen binding probability based on dilated convolution and channel attention mechanism according to claim 1, characterized in that, In the feature extraction step, the dilated convolution module contains several parallel convolutional branches with different dilation rates, which are used to extract local and global features of the sequence within different receptive fields. In the feature extraction step, the number of channels in each group of dilated convolution modules increases layer by layer to enhance the feature representation capability. In the feature extraction step, residual connections and random deactivation are used at the output of each dilated convolution module to enhance feature transfer and prevent overfitting.

3. The method for predicting antibody-antigen binding probability based on dilated convolution and channel attention mechanism according to claim 1, characterized in that, The steps of the channel attention mechanism weighting and feature aggregation include: performing global pooling on the features of each channel in the matrix processed by the feature extraction step to obtain a channel description vector; then mapping it through two fully connected networks to output a channel weight vector; and multiplying the weight vector with the original features channel by channel to highlight key features, suppress redundant features, and improve feature selectivity. In the two-layer fully connected network, the first layer is activated using a linear rectified function, and the second layer is activated using a sigmoid function.

4. The method for predicting antibody-antigen binding probability based on dilated convolution and channel attention mechanism according to claim 1, characterized in that, In the classification prediction step, the fused features in the fused feature matrix are input into the classification prediction module for classification processing; The classification prediction module is implemented through a fully connected classification module, which includes a multi-layer feature mapping unit and sets a residual connection structure between each layer to enhance feature transfer and model stability. The feature mapping unit is equipped with a nonlinear activation and regularization mechanism to improve the nonlinear expression ability of the model and reduce the risk of overfitting; finally, the model output is equipped with a binding probability prediction unit to output the binding probability between the antibody and the antigen.

5. The method for predicting antibody-antigen binding probability based on dilated convolution and channel attention mechanism according to claim 1, characterized in that, In the model training and optimization steps, binary cross-entropy is used as the loss function, and the difference between the predicted probability value and the true label is used as the optimization objective.

6. A predictive model for antibody-antigen binding probability based on dilated convolution and channel attention mechanisms, characterized in that, The prediction model includes: Data preprocessing module: acquires the heavy chain amino acid sequence of the antibody, the light chain amino acid sequence of the antigen, encodes the amino acid sequence, and converts the amino acid sequence into a fixed-length one-hot feature matrix; Feature extraction module: The one-hot feature matrix is ​​input into the encoder, which is composed of several dilated convolution modules stacked sequentially; the matrix output after dilation convolution uses residual connection and random deactivation to enhance feature transfer and prevent overfitting. Attention weighting and feature aggregation module: Apply channel attention mechanism to the matrix processed by the feature extraction module to perform attention weighting to enhance feature representation and output channel weight vector. Multiply the weight vector with the original features channel by channel. Global pooling is performed on the feature maps output by the heavy chain, light chain and antigen coding branches after attention weighting to obtain fixed-length feature vectors. Then, the three are concatenated to form a fusion feature representation, forming a unified fusion feature matrix to provide input for subsequent classification prediction. Classification and prediction module: performs classification and prediction on the fused feature matrix and outputs a predicted value of the probability of antibody-antigen binding; Training and Optimization Module: Acquires positive and negative sample data as training and validation sets, respectively. Positive sample data refers to antibody sequences that can bind to antigens, while negative sample data refers to antibody sequences that cannot bind to antigens. Based on the labeled positive and negative sample data, training is performed using the cross-entropy loss function and an optimizer. The model parameters are saved when the validation accuracy and loss rate meet preset conditions. The preset conditions are that when the validation accuracy is greater than or equal to the historical best value and the validation loss is less than or equal to the historical minimum value, the current model parameters are saved as the optimal model parameters. One-hot encoding is used to digitally represent amino acid sequences, where each amino acid is converted into a one-hot vector of length 20, corresponding to 20 standard amino acids. To ensure consistent input dimensions, sequences that are too short are padded to a uniform length. After encoding and padding, a single sequence is represented as a two-dimensional feature matrix composed of feature vectors of multiple amino acids. Multiple encoded and padded sequences can be combined in batches to form the model input features for subsequent feature extraction and learning.

7. The antibody-antigen binding probability prediction model based on dilated convolution and channel attention mechanism according to claim 6, characterized in that, In the data preprocessing module In the feature extraction module, the dilated convolution module contains several parallel convolutional branches with different dilation rates, which are used to extract local and global features of the sequence within different receptive fields. In the feature extraction module, the number of channels in each group of dilated convolution modules increases layer by layer to enhance the feature representation capability. In the feature extraction module, the output of each dilated convolution module uses residual connection and random deactivation to enhance feature transfer and prevent overfitting; The steps of the channel attention mechanism weighting and feature aggregation include: performing global pooling on the features of each channel in the matrix processed by the feature extraction step to obtain a channel description vector; then mapping it through two fully connected networks to output a channel weight vector; and multiplying the weight vector with the original features channel by channel to highlight key features, suppress redundant features, and improve feature selectivity. In the two-layer fully connected network, the first layer is activated using a linear rectified function, and the second layer is activated using a sigmoid function. In the classification prediction module, the fused features in the fused feature matrix are input to the special classification prediction module for classification processing; The classification prediction module is implemented through a fully connected classification module, which includes a multi-layer feature mapping unit and sets a residual connection structure between each layer to enhance feature transfer and model stability. The feature mapping unit is equipped with a nonlinear activation and regularization mechanism to improve the nonlinear expression ability of the model and reduce the risk of overfitting; finally, a binding probability prediction unit is set at the model output end to output the binding probability between the antibody and the antigen. In the training and optimization module, binary cross-entropy is used as the loss function, and the difference between the predicted probability value and the true label is used as the optimization objective.

8. The method for predicting the probability of antibody-antigen binding based on dilated convolution and channel attention mechanism as described in any one of claims 1-5, or the application of the prediction model for antibody-antigen binding probability based on dilated convolution and channel attention mechanism as described in claim 6 or 7 in antibody screening.

9. A computer device comprising a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method for predicting the probability of antibody-antigen binding based on dilated convolution and channel attention mechanism as described in any one of claims 1-5.

Images