An antibody structure generation method based on deep learning

Abstract

The application belongs to the technical field of biological information, and particularly relates to an antibody structure generation method based on deep learning, which comprises the following steps: obtaining sequence data to be predicted, and pre-processing the data; adopting a PCA method to perform decomposition dimension reduction processing on the pre-processed feature sequence, the decomposition dimension reduction processing process comprising calculating a covariance matrix of sequence features, and calculating eigenvalues of the covariance matrix; sorting the eigenvalues, screening corresponding eigenvectors according to the sorted eigenvalues, and taking the screened eigenvectors as sequence features after decomposition dimension reduction; inputting the sequence data after decomposition dimension reduction into an improved neural network model trained, and obtaining an antibody structure prediction result; and the application can more accurately express the spatial structure of the antibody by extracting all heavy atom information of each amino acid as a label for the antibody structure.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a method for generating antibody structures based on deep learning. Background Technology

[0002] Antibodies are proteins that play a crucial role in humoral immunity. Their structure consists of two heavy (H) and two light (L) polypeptide chains. Their primary function is to bind to invading foreign organisms with high affinity and specificity. Antibodies are produced to mediate an immune response against foreign pathogens as part of adaptive immunity. With the development of artificial intelligence (AI) technology, especially deep learning (DL), and the accumulation of antibody structure data, deep learning-based antibody structure prediction has made significant progress. In drug discovery and protein engineering, a major goal is to design antibodies that can function as therapeutic drugs. This requires the analysis of antibody structures. Traditional methods for antibody analysis are time-consuming, labor-intensive, and very expensive. Driven by advancements in computer vision and natural language processing, deep learning technology has been widely applied in the antibody field in recent years.

[0003] An antibody monomer consists of two heavy chains (H) and two light chains (L), which are linked together to form a "Y"-shaped structure. It has an NH2-terminal variable region, or antigen-binding fragment (Fab), and a COOH constant region, or crystallizable fragment (Fc). Different parts of the structure have different functions. The variable region determines the antibody's idiotype and has affinity for pathogen antigens. The constant region performs other immune-related functions, such as complement binding and macrophage binding; in studies, the antibody can be labeled with isotopes.

[0004] Antibody antigen-binding sites are located in the Fab region, primarily composed of six complementarity-determining regions (CDRs): three each for the heavy and light chains. Due to the diversity generated during antibody synthesis, antibodies can recognize millions of different antigens. The heavy and light chains undergo genetic recombination, resulting in different gene combinations to increase the diversity of the final gene product. Of the six antibody CDRs, five typically fold into one of several classic conformations, which can be predicted using existing methods. However, the remaining loop (CDR H3), due to its highly variable experimentally observed structure, cannot be predicted well by conventional methods. To perform computational analysis of antibodies or predict their effectiveness, a three-dimensional model is usually required. However, traditional structural determination methods such as X-ray crystallography, nuclear magnetic resonance (NMR), and cryo-electron microscopy (CryoEM) are laborious, time-consuming, and expensive.

[0005] Existing machine learning methods for protein structure prediction focus on co-evolutionary approaches. The accuracy of these methods depends on the number of homologous protein sequences available in the database. For many proteins, especially those lacking sufficient sequence homologs, such as antibodies, co-evolutionary methods perform poorly, resulting in low accuracy in predicted structures. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention proposes a deep learning-based antibody structure generation method, comprising: acquiring sequence data to be predicted and preprocessing the data; inputting the preprocessed sequence data into a trained improved neural network model to obtain the predicted antibody structure; constructing an antibody based on the predicted antibody structure; and modifying the improved neural network model to an improved ResNet-RCCA model.

[0007] The process of training an improved neural network model includes:

[0008] S1: Obtain the raw antibody dataset, which contains the sequence information of the antibody, including the primary structure, secondary structure and tertiary structure of the antibody;

[0009] S2: Convert the antibody sequence information in the original antibody dataset into matrix data and use the matrix data as sequence features; obtain the atomic coordinate information of the antibody structure and use this information as structural labels;

[0010] S3: The PCA method is used to decompose and reduce the dimensionality of the sequence features;

[0011] S4: Input the structural labels and the decomposed and dimensionality-reduced sequence features into the improved ResNet-RCCA model to obtain the antibody structure prediction results;

[0012] S5: Calculate the model's loss function based on the antibody structure prediction results, continuously adjust the model parameters, and complete model training when the loss function is minimized.

[0013] Preferably, the process of converting antibody sequence information in the original antibody dataset into matrix data includes: inputting the antibody sequence information into a pre-trained protein semantic model ESM-1B to obtain matrix data with antibody sequence information, wherein the protein semantic model ESM-1B is a high-capacity Transformer model that takes protein sequences as input and is trained through hyperparameter optimization.

[0014] Preferably, the process of using PCA to decompose and reduce the dimensionality of sequence features includes: calculating the covariance matrix of the sequence features and calculating the eigenvalues ​​of the covariance matrix; sorting the eigenvalues, selecting the corresponding eigenvectors based on the sorted eigenvalues, and using the selected features as the sequence features after decomposition and dimensionality reduction.

[0015] Preferably, the improved ResNet-RCCA model includes a ResNet model and 6 RCCA modules. The ResNet model consists of a one-dimensional residual convolutional network and a two-dimensional residual convolutional network. Each one-dimensional convolutional block consists of two convolutional layers and two pooling layers, with a convolutional kernel size of 5*5. The two-dimensional residual convolutional network consists of one two-dimensional convolutional layer, a pooling layer, and 25 two-dimensional convolutional blocks. Each two-dimensional convolutional block consists of two convolutional layers and two pooling layers, with a convolutional kernel size of 3*3. The RCCA modules are placed after the two-dimensional convolutional layers.

[0016] Preferably, the process of processing the structural labels and decomposed / dimensionalized sequence features using the improved ResNet-RCCA model includes: inputting the decomposed / dimensionalized sequence features into a one-dimensional residual convolutional network to obtain a one-dimensional feature map; concatenating the one-dimensional feature map with the input sequence features, and inputting the concatenated feature map into a two-dimensional residual convolutional network to obtain a two-dimensional feature map; inputting the two-dimensional feature map into the RCCA module, learning the temporal information of the two-dimensional feature map through the residual convolutional network, and learning the spatial information of the two-dimensional feature map using a cross-attention mechanism; and predicting the antibody structure based on the temporal and spatial information of the two-dimensional feature map to obtain the predicted antibody structure.

[0017] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program is executed by a processor, and when the computer program is executed by the processor, it implements any of the above-described deep learning-based antibody structure generation methods.

[0018] To achieve the above objectives, the present invention also provides a deep learning-based antibody structure generation device, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the deep learning-based antibody structure generation device performs any of the above-described deep learning-based antibody structure generation methods.

[0019] The present invention has the following beneficial effects:

[0020] This invention fully utilizes antibody sequence information and amino acid feature information, without applying co-evolutionary information to predict antibody structure; for antibody structure, this invention extracts all heavy atom information for each amino acid as a tag, which can more accurately express the spatial structure of the antibody; this invention proposes a novel hybrid neural network that integrates ResNet and cross-attention mechanism. The fused network can more effectively extract and learn protein sequence information, improve prediction accuracy, and has good generalization performance for different datasets. Attached Figure Description

[0021] Figure 1 This is a flowchart of the antibody structure prediction method based on deep learning of the present invention;

[0022] Figure 2 This is a structural diagram of the model framework of the present invention;

[0023] Figure 3 This is a diagram of the improved neural network structure of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] A deep learning-based method for generating antibody structures, such as Figures 1-2 As shown, the method includes: acquiring the sequence data to be predicted and preprocessing the data; inputting the preprocessed sequence data into a trained improved neural network model to obtain the predicted antibody structure; constructing the antibody based on the predicted antibody structure; and improving the neural network model to obtain the improved ResNet-RCCA model.

[0026] The process of training an improved neural network model includes:

[0027] S1: Obtain the raw antibody dataset, which contains the sequence information of the antibody, including the primary structure, secondary structure and tertiary structure of the antibody;

[0028] S2: Convert the antibody sequence information in the original antibody dataset into matrix data and use the matrix data as sequence features; obtain the atomic coordinate information of the antibody structure and use this information as structural labels;

[0029] S3: The PCA method is used to decompose and reduce the dimensionality of the sequence features;

[0030] S4: Input the structural labels and the decomposed and dimensionality-reduced sequence features into the improved ResNet-RCCA model to obtain the antibody structure prediction results;

[0031] S5: Calculate the model's loss function based on the antibody structure prediction results, continuously adjust the model parameters, and complete model training when the loss function is minimized.

[0032] The process of using PCA to decompose and reduce the dimensionality of sequence features includes: calculating the covariance matrix of the sequence features, performing eigenvalue decomposition on the covariance matrix to obtain its eigenvalues ​​and eigenvectors, and sorting the eigenvalues ​​from largest to smallest; selecting the corresponding eigenvectors based on the sorted eigenvalues, and using the selected features as the sequence features after dimensionality reduction. Selecting the top 100 eigenvectors involves selecting the corresponding eigenvectors.

[0033] Antibody primary structure data is transformed into matrix data using a protein semantic model as sequence features, and heavy atom coordinate information (Cα atom, Cβ atom, N atom, O atom) is extracted from antibody tertiary structure data as structural tags.

[0034] In this example, the SAbdab database is used, which contains corresponding audio and image data. The process of transforming the data in the training set includes:

[0035] Antibody structure data includes information on the primary, secondary, and tertiary structures of antibodies. The antibody structure data is processed to extract the primary and tertiary structure information. For the primary structure, a pre-trained ESM-1b model is used to transform it from sequence data into matrix data. Dimensionality reduction is performed using the PAC method, and then the matrix data is transposed to fix its dimensions.

[0036] For each amino acid in the antibody's tertiary structure, the coordinate information of heavy atoms (Cα, Cβ, N, and O atoms) is extracted as structural tags to represent the antibody's tertiary structure.

[0037] The pre-trained protein semantic model ESM-1b was used. ESM-1b is actually a high-capacity Transformer trained by taking protein sequences as input and optimizing them with hyperparameters. After training, the feature representation output by this model implicitly contains information such as the secondary and tertiary structure, function, and homology of proteins, and this information can be made explicit through linear projection.

[0038] like Figure 3As shown, the improved ResNets-RCCA model includes the ResNet model and six RCCA modules. The ResNet model consists of a one-dimensional residual convolutional network and a two-dimensional residual convolutional network. The one-dimensional residual convolutional network consists of one one-dimensional convolutional layer, a pooling layer, and three one-dimensional convolutional blocks. Each one-dimensional convolutional block consists of two convolutional layers and two pooling layers, with a kernel size of 5*5 for each convolutional layer. The two-dimensional residual convolutional network consists of one two-dimensional convolutional layer, a pooling layer, and 25 two-dimensional convolutional blocks. Each two-dimensional convolutional block consists of two convolutional layers and two pooling layers, with a kernel size of 3*3 for each convolutional layer. The RCCA modules are placed after the two-dimensional convolutional layers.

[0039] The improved ResNets-RCCA model includes a ResNet model and six RCCA modules: it processes antibody primary structure data through a protein language model to obtain sequence information as feature input, uses residual convolutional networks and cross-attention mechanisms to learn temporal information, predicts antibody structure, and finally obtains the prediction result.

[0040] The process of processing structural labels and decomposed / dimensionalized sequence features using the improved ResNet-RCCA model includes: inputting the decomposed / dimensionalized sequence features into a one-dimensional residual convolutional network to obtain a one-dimensional feature map; concatenating the one-dimensional feature map with the input sequence features and inputting the concatenated feature map into a two-dimensional residual convolutional network to obtain a two-dimensional feature map; inputting the two-dimensional feature map into the RCCA module, learning the temporal information of the two-dimensional feature map through the residual convolutional network, and learning the spatial information of the two-dimensional feature map using a cross-attention mechanism; and predicting the antibody structure based on the temporal and spatial information of the two-dimensional feature map to obtain the predicted antibody structure.

[0041] When constructing a neural network, the process of calculating the model's loss function includes performing a cross-entropy calculation between the joint vector features of the input and the actual label of that vector, and using the cross-entropy as the model's loss function; the expression for the model's loss function is:

[0042]

[0043] Where, y′ i For the actual label; y i ...

[0044] In one embodiment of the present invention, the present invention further includes a computer-readable storage medium having a computer program stored thereon, wherein the method, when executed by a processor, implements any of the deep learning antibody structure generation methods described above.

[0045] A deep learning-based antibody structure generation device includes a processor and a memory; the memory stores a computer program; the processor is connected to the memory and executes the computer program stored in the memory, causing the deep learning-based antibody structure generation device to perform any of the aforementioned deep learning-based antibody structure generation methods.

[0046] Specifically, the memory includes various media capable of storing program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card, or optical disk.

[0047] Preferably, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0048] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating antibody structures based on deep learning, characterized in that, include: Obtain the sequence data to be predicted and preprocess the data; The preprocessed sequence data is input into a trained improved neural network model to obtain the predicted antibody structure; the antibody is then constructed based on the predicted antibody structure. An improved neural network model: the ResNet-RCCA model; The process of training an improved neural network model includes: S1: Obtain the raw antibody dataset, which contains the sequence information of the antibody, including the primary structure, secondary structure and tertiary structure of the antibody; S2: Convert the antibody sequence information in the original antibody dataset into matrix data and use the matrix data as sequence features; obtain the atomic coordinate information of the antibody structure and use this information as structural labels; S3: The PCA method is used to decompose and reduce the dimensionality of the sequence features; S4: Input the structural labels and the decomposed and dimensionality-reduced sequence features into the improved ResNet-RCCA model to obtain the antibody structure prediction results; The improved ResNet-RCCA model includes a ResNet model and six RCCA modules. The ResNet model consists of a one-dimensional residual convolutional network and a two-dimensional residual convolutional network. Each one-dimensional convolutional block consists of two convolutional layers and two pooling layers, with a kernel size of 5*5 for each convolutional layer. The two-dimensional residual convolutional network consists of one two-dimensional convolutional layer, a pooling layer, and 25 two-dimensional convolutional blocks. Each two-dimensional convolutional block consists of two convolutional layers and two pooling layers, with a kernel size of 3*3 for each convolutional layer. The RCCA modules are placed after the two-dimensional convolutional layers. S5: Calculate the model's loss function based on the antibody structure prediction results, continuously adjust the model parameters, and complete model training when the loss function is minimized.

2. The method for generating antibody structures based on deep learning according to claim 1, characterized in that, The process of converting antibody sequence information in the original antibody dataset into matrix data includes: inputting the antibody sequence information into the pre-trained protein semantic model ESM-1B to obtain matrix data with antibody sequence information. The protein semantic model ESM-1B is a high-capacity Transformer model that takes protein sequences as input and is trained through hyperparameter optimization.

3. The method for generating antibody structures based on deep learning according to claim 1, characterized in that, The process of using PCA to decompose and reduce the dimensionality of sequence features includes: calculating the covariance matrix of the sequence features, performing eigenvalue decomposition on the covariance matrix to obtain its eigenvalues ​​and eigenvectors, and sorting the eigenvalues ​​from largest to smallest; selecting the corresponding eigenvectors based on the sorted eigenvalues, and using the selected features as the sequence features after dimensionality reduction.

4. The method for generating antibody structures based on deep learning according to claim 1, characterized in that, The process of processing structural labels and decomposed / dimensionalized sequence features using the improved ResNet-RCCA model includes: inputting the decomposed / dimensionalized sequence features into a one-dimensional residual convolutional network to obtain a one-dimensional feature map; concatenating the one-dimensional feature map with the input sequence features and inputting the concatenated feature map into a two-dimensional residual convolutional network to obtain a two-dimensional feature map; inputting the two-dimensional feature map into the RCCA module, learning the temporal information of the two-dimensional feature map through the residual convolutional network, and learning the spatial information of the two-dimensional feature map using a cross-attention mechanism; and predicting the antibody structure based on the temporal and spatial information of the two-dimensional feature map to obtain the predicted antibody structure.

5. The method for generating antibody structures based on deep learning according to claim 1, characterized in that, The loss function expression for the model is: ； in, For actual labels; ...

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor to implement any one of the deep learning-based antibody structure generation methods of claims 1 to 5.

7. A deep learning-based antibody structure generation device, characterized in that, It includes a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the deep learning-based antibody structure generation device performs any one of the deep learning-based antibody structure generation methods of claims 1 to 5.

Images