An Antibody-Oriented Prediction Method for Double-Branched Structural Features

Abstract

The present invention discloses a method for predicting the dual-branch structural features of antibodies, which relates to the technical fields of protein representation learning and protein structure prediction, and includes the following steps: Step 1, for a given antibody amino acid sequence as input, extract three types of features of the antibody, namely MSA features, Pair features, and template features, and perform collaborative fusion on the three types of features; Step 2, construct a dual-branch structural feature prediction module to optimize the antibody structure; Step 3, model training; Step 4, model testing to obtain the antibody structure. By adopting the above method for predicting the dual-branch structural features of antibodies, the present invention realizes efficient and accurate prediction of antibody structural features, providing strong support for antibody-related research and applications.

Description

Technical Field

[0001] The present invention relates to the technical fields of protein representation learning and protein structure prediction, and in particular to a dual-branch structural feature prediction method for antibodies. Background Art

[0002] Antibody structure prediction, that is, in the case of knowing the amino acid sequence of an antibody, predicting the three-dimensional spatial structure of the antibody by computational methods. Antibodies are a special type of protein produced by the immune response of an organism to invading pathogens. Their main function is to recognize antigens and specifically bind to them, thereby effectively clearing foreign substances such as bacteria and fungi invading the human body, neutralizing the toxins released by these pathogens, and maintaining the balance of the immune system, which plays a significant role in maintaining human health. As a key component of the immune system, the function of an antibody is closely related to its structure. Therefore, accurate modeling and prediction of antibody structures are of great significance for understanding the antibody-antigen interaction mechanism, and even for drug design, vaccine development, and disease treatment.

[0003] The variable region of an antibody is the key structural basis for its ability to bind various antigens. The variable region consists of a framework region (FR) and complementarity-determining regions (CDR). The framework region provides the basic structural support, while the high variability of the complementarity-determining regions, especially the heavy-chain complementarity-determining region 3 (H3), enables the antibody to adapt to the structures of different antigens. The complementarity-determining regions can form loop structures of various shapes, and these structures can be complementary to the epitopes of antigens in space. This extreme diversity makes it difficult for general methods to effectively predict their structures, and their importance requires the prediction results to be highly accurate.

[0004] There have been various methods for the research of antibody structure prediction tasks:

[0005] Traditional methods: Traditional methods can basically be divided into homology modeling methods and de novo prediction methods, mainly relying on experimental means such as X-ray crystallography and nuclear magnetic resonance imaging. These methods are not only time-consuming but also costly.

[0006] Artificial intelligence-based methods: Deep learning builds a neural network model and uses a large amount of known antibody structure and sequence data for training. It can learn complex patterns and features from a large amount of data, thereby achieving high-precision prediction of antibody structures. However, the prediction of antibody structures not only needs to consider the information of amino acid sequences but also needs to comprehensively consider their spatial conformations, that is, three-dimensional structures. Traditional sequence-based deep learning methods are difficult to directly process such structured data, which limits their application effects in antibody structure prediction. Summary of the Invention

[0007] The object of the present invention is to provide a method for predicting the dual-branch structural features of antibodies. By extracting multi-dimensional features from antibodies and aggregating and updating them, the pairwise features are used to capture the complex relationships within the antibodies, and the structural features of the antibodies are predicted to assist in the task, so as to achieve the purpose of more accurate prediction of the three-dimensional structure of the antibodies.

[0008] To achieve the above object, the present invention provides a method for predicting the dual-branch structural features of antibodies, including the following steps:

[0009] Step 1: For a given antibody amino acid sequence as input, three types of features, namely MSA features, Pair features, and template features of the antibody, are extracted by combining templates, and the three types of features are synergistically fused;

[0010] Step 2: Construct a dual-branch structural feature prediction module to optimize the antibody structure;

[0011] Step 3: Model training;

[0012] Step 4: Model testing to obtain the antibody structure.

[0013] Preferably, in Step 2, the target antibody structural features are represented as a set of six relative distance and direction matrices in the residue pair L×L, including three inter-residue distance matrices and three inter-residue direction matrices.

[0014] Preferably, in Step 2, a structural mode analysis network based on residue cross-attention is constructed for prediction for the three inter-residue distance matrices and the two inter-residue direction matrices θ and φ, including the following steps:

[0015] S11: Add the Pair feature to its transposed matrix and take the average to obtain a symmetric matrix as the input;

[0016] S12: Pass through a module stacked by 3 residual blocks, and use the residual connection to add the processed feature to the original input and then activate the output to perform deep feature extraction and transformation;

[0017] S13: After passing through a convolutional layer to project the dimension to the dimension corresponding to the set range interval number, the output feature passes through two layers of cross-attention in sequence to mine the correlation information of the features between residue pairs;

[0018] S14: Process the feature again through an operation sequence composed of a convolutional layer with a convolution kernel of 3*3, batch normalization, and ReLU activation;

[0019] S15: Adjust the dimension of the feature through a linear transformation layer and a transpose operation, and output the distance and direction matrices respectively.

[0020] Preferably, in S13, dCA , d CB , d NO , θ is projected to 37, and φ is projected to 18.

[0021] Preferably, for the inter-residue direction matrix, using the Pair feature as the input, a structure symmetry state deduction network guided by dilated convolution is constructed for prediction, including the following steps:

[0022] S21. Add the Pair feature and its transposed matrix and take the average to obtain a symmetric matrix as the input;

[0023] S22. Through multiple convolution modules, including a convolutional layer, an instance normalization layer, and an ELU activation function layer, add the features output by multiple layers of convolution to the original input and perform ELU activation to achieve residual connection;

[0024] S23. Adjust the dimension through a convolutional layer and a linear transformation layer, and after the transpose operation, output a dihedral angle ω direction matrix with a shape of 1, nres, nres, 37, where nres is the number of amino acid residues in the antibody sequence.

[0025] Preferably, in S22, the convolutional layer sequentially uses convolutional kernels with increasing sizes and dilation rates for dilated convolution to expand the receptive field.

[0026] Preferably, in the model training of step three, the RoseTTAFold2 network is used to extract features from the antibody sequence.

[0027] Preferably, in the model training of step three, the activation functions used are the exponential linear unit ELU and the rectified linear unit ReLU, the dropout is set to 0.1 to avoid overfitting, the optimizer is Adam, trained for 10 epochs, the learning rate is set to 0.001, and the learning rate scheduler MultiStepLR is used to reduce the learning rate by 0.5 times at the 2nd, 3rd, 4th, 6th, and 8th epochs.

[0028] Therefore, the present invention adopts the above-mentioned antibody-oriented double-branch structure feature prediction method, which realizes efficient and accurate antibody structure feature prediction and provides strong support for antibody-related research and applications. By constructing a double-branch structure feature prediction module and using a structure mode analysis network based on residue cross-attention, the d CA distance matrix, d CB distance matrix, d NOFive structural features such as the distance matrix, the dihedral angle θ direction matrix, and the torsion angle φ direction matrix; using a structure symmetry state deduction network guided by dilated convolution to separately predict and output one structural feature, the dihedral angle ω direction matrix. These six antibody structural features can provide necessary constraints for antibody structure prediction training, making the predicted structure more in line with biological reality, thereby effectively improving the accuracy and reliability of the task and enhancing the adaptability and effectiveness of the overall prediction effect in biological application scenarios.

[0029] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Brief Description of the Drawings

[0030] Figure 1 It is a schematic structural diagram of an embodiment of a dual-branch structural feature prediction method for antibodies according to the present invention;

[0031] Figure 2 It is a front view of an embodiment of a dual-branch structural feature prediction method for antibodies according to the present invention. Detailed Embodiments

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left", "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0034] Embodiment

[0035] Please refer to Figure 1-2, the present invention provides a method for predicting the dual-branch structure characteristics of antibodies. Based on the input antibody amino acid sequence, homologous sequences are searched, and multiple sequence alignment (MSA) characteristics, pairwise (Pair) characteristics, and template characteristics of the sequence are obtained through feature extraction. The three types of characteristics are interactively updated to promote the full integration of information between different characteristics, enabling the characteristics to more accurately reflect the structural and functional relationship of antibodies. A dual-branch structure characteristic prediction module is constructed to process the Pair characteristics separately. Different prediction networks in the two branches are used to predict six relative distance and direction matrices between residue pairs in the sequence, which can provide assistance for the training process of antibody structure prediction to optimize the final predicted structure.

[0036] Step 1: For a given antibody amino acid sequence as input, use a specific tool to extract the msa file containing its multiple sequence alignment content, and extract three types of characteristics of the antibody, namely MSA characteristics, Pair characteristics, and template characteristics, in combination with the template. The three types of characteristics are synergistically fused to learn feature information and potential associations with each other, and an initial antibody prediction structure is obtained.

[0037] Step 2: The dual-branch structure characteristic prediction module has one branch as a structural modal analysis network based on residue cross-attention, and one branch as a structural symmetry state deduction network guided by dilated convolution. The three types of characteristics obtained by feature extraction of the antibody are multiple sequence alignment MSA characteristics, pairwise Pair characteristics, and template characteristics, which respectively carry key information in different dimensions of the antibody, such as one-dimensional gene sequence, two-dimensional folding rotation, and three-dimensional spatial conformation. The three are fused and updated to learn feature information and potential associations with each other.

[0038] Considering the huge differences and instabilities in the variable region structures of different antibodies, the prediction of structural characteristics can play a certain auxiliary role in antibody structure prediction. For a target antibody sequence with L residues, the target antibody structural characteristics can be represented as a set of six relative distance and direction matrices in the residue pairs (L×L), including three residue distance matrices: C α between atoms (d CA : C α -C α ), C β between atoms (d CB : C β -C β ), between N-O atoms (d NO : N-O); and three residue direction matrices: two dihedral angles (ω: C α -C β -C β -C α , θ: N-C α -C β -Cβ ), a torsional angle (φ: C α -C β -C β ). Classify each specific value in the six matrices and map them to different intervals: For the three distance matrices, divide the residue-residue distances into 37 intervals. The width of the first 36 intervals is representing the length range, and the 37th interval represents that the residue-residue distance exceeds For the two dihedral angle matrices, divide the residue-residue dihedral angles into 36 intervals with a width of 10°, representing the angular range of 0 - 360°; for one torsional angle matrix, divide the residue-residue torsional angles into 18 intervals with a width of 10°, representing the angular range of 0 - 180°.

[0039] Construct a dual-branch structure feature prediction module to predict the relative distance and orientation matrix between residues. Use the Pair feature as the module input and the relative distance and orientation matrix between residues as the module output. Construct a structure mode analysis network based on residue cross-attention and a structure symmetry state deduction network guided by dilated convolution to perform dual-branch prediction on the relative distance and orientation matrix between residues. The following are the specific network architectures of the two branches:

[0040] For the three residue-residue distance matrices, dihedral angle θ, and torsional angle φ orientation matrices, use the Pair feature as the input and construct a structure mode analysis network based on residue cross-attention for prediction. Among them, due to the symmetry property of d CA , d CB , add the Pair feature to its transposed matrix and take the average to obtain a symmetric matrix as the input. The feature passes through a module stacked by 3 residual blocks in sequence. Inside each residual block, first ensure that the convolution size remains the same before and after convolution through appropriate padding, and then include operations such as convolution, instance normalization, ELU activation, Dropout, etc. Moreover, use the residual connection to add the processed feature to the original input and then activate the output to perform in-depth feature extraction and transformation. Then, pass through a convolutional layer to project the dimension to the dimension corresponding to the set range interval number (where d CA , d CB , d NO, θ is projected to 37, and φ is projected to 18). Subsequently, the output features pass through two layers of CrossCrossAttention successively. Through its unique query, key, value convolutions, dimension transformation, attention calculation and other operations, the correlation information between residue pairs is further mined, so that each output residue pair fully aggregates the information of all other residue pairs. Then, the features are processed again by an operation sequence consisting of a convolutional layer with a convolution kernel of 3*3, batch normalization, and ReLU activation. Finally, a linear transformation layer and a transpose operation are used to adjust the dimensions of the features, and distance and direction matrices with shapes of [1, nres, nres, 37] (where the shape of φ is [1, nres, nres, 18]) are output respectively, where nres is the number of amino acid residues in the antibody sequence.

[0041] For the more complex symmetric dihedral angle ω direction matrix, taking the Pair feature as the input, a structure symmetry state deduction network guided by dilated convolution is constructed for prediction. Due to the symmetric property of ω, the Pair feature is added to its transposed matrix and averaged to obtain a symmetric matrix as the input. The feature passes through multiple convolutional modules, each of which includes a convolutional layer, an instance normalization layer, and an ELU activation function layer. The features output by multiple layers of convolution are added to the original input and activated by ELU to implement residual connection. The convolutional layers in it use dilated convolution with successively increasing kernel sizes (3*3, 5*5, etc.) and dilation rates (1, 2, 4, 8, etc.) in sequence to expand the receptive field, so that the symmetric Pair features in the antibody structure can better capture the feature relationships at different distance scales. For example, the symmetric relationships between amino acid residues at close and far distances are considered simultaneously, so as to more comprehensively extract symmetric features. Finally, the feature passes through a convolutional layer and a linear transformation layer to adjust the dimensions, and after the transpose operation, a dihedral angle ω direction matrix with a shape of [1, nres, nres, 37] is output, where nres is the number of amino acid residues in the antibody sequence.

[0042] After predicting six relative distance and direction matrices, the cross-entropy loss is calculated with the real structure to optimize the network.

[0043] Step 3. Model training stage

[0044] 3.1 Training dataset

[0045] To construct the dataset used in the training process, we selected all paired structural antibodies collected before July 2021 from the structural antibody dataset website SAbDab, excluding antibody data with missing FV, incomplete heavy and light chains in the FV region, and missing heavy and light chains in the FV region, and screened 2,640 antibody samples. The heavy and light chains of these antibodies were separated, and the Chothia sequences of these single chains were taken. The msa files were obtained through the HHb l its tool as the model training dataset.

[0046] 3.2 Training Process

[0047] Perform training on the training dataset according to the steps in Section 2. The specific details are as follows:

[0048] The RoseTTAFol d2 network was used to extract features from the antibody sequence, and the MSA features were 256-dimensional, the Pair features were 128-dimensional, and the template features were 64-dimensional. The number of cross-attention heads in the residue-based cross-attention structural modal parsing network was set to 8. The activation function used exponential linear unit (ELU) and rectified linear unit (ReLU), the dropout was set to 0.1 to avoid overfitting, the optimizer used Adam, and the training was 10 cycles. The learning rate was set to 0.001, and the learning rate scheduler Mult iStepLR was used to reduce the learning rate by 0.5 times in the 2nd, 3rd, 4th, 6th, and 8th cycles.

[0049] 3.3 Training Loss Function

[0050] We use a combination of multiple loss functions to construct the overall loss function of the training process. The specific contents are as follows:

[0051] L=λ1L CA +λ2L CB +λ3L NO +λ4L ω +λ5L θ +λ6L φ ;

[0052] Among them, L CA ,L CB ,L NO ,L ω ,L θ ,L φ They are d CA ,d CB ,d NO ,ω,θ,φ are the losses, λ1~λ6 are the weight parameters corresponding to each loss, so as to adjust each loss to a suitable order of magnitude. All losses are calculated in the form of cross entropy loss.

[0053] Finally, the training is completed and the model is obtained.

[0054] Step Four: Model Testing Phase

[0055] In the testing phase, only the model obtained by the above training steps is needed. The antibody sequences (fasta files) of the test dataset are input into the model to obtain the predicted antibody structure features and the final antibody structure of the model.

[0056] Therefore, the present invention adopts the above-mentioned method for predicting double-branch structural features of antibodies, realizes efficient and accurate prediction of antibody structural features, and provides strong support for antibody-related research and applications. By constructing a double-branch structural feature prediction module and using a structural mode parsing network based on residue cross-attention, five structural features such as the d CA distance matrix, d CB distance matrix, d NO distance matrix, dihedral angle θ direction matrix, and torsion angle φ direction matrix are predicted and output; a structural symmetry state deduction network guided by dilated convolution is used to separately predict and output one structural feature, namely the dihedral angle ω direction matrix. These six antibody structural features can provide necessary constraint conditions for antibody structure prediction training, making the predicted structure more in line with biological reality, thereby effectively improving the accuracy and reliability of the task and enhancing the adaptability and effectiveness of the overall prediction effect in biological application scenarios.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements do not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An antibody-oriented method for predicting the characteristics of a double-branched structure, characterized in that, It includes the following steps: Step 1: For a given antibody amino acid sequence as input, extract three types of features, namely MSA features, Pair features, and template features of the antibody from the template, and perform collaborative fusion on the three types of features; Step 2: Construct a dual-branch structure feature prediction module to optimize the antibody structure; Step 3: Model training; Step 4: Model testing to obtain the antibody structure; In step two, the structural features of the target antibody are represented as residue pairs A set of six relative distance and direction matrices in total, including three inter-residue distance matrices and three inter-residue direction matrices; In step two, for the distance matrix between three residues and the direction matrix between two residues and Construct a structure mode parsing network based on residue cross-attention for prediction, including the following steps: S11: Add the Pair feature to its transpose matrix and take the average to obtain a symmetric matrix as the input; S12: Pass through a module stacked by 3 residual blocks, use residual connections to add the processed features to the original input and then activate the output to perform deep feature extraction and transformation; S13: Project the dimension to the dimension corresponding to the set range interval number through a convolutional layer, and the output features successively pass through two layers of cross-attention to mine the correlation information between residue pairs; S14. Process the features again through an operation sequence consisting of a convolutional layer with a convolutional kernel of , batch normalization, and ReLU activation; S15: Adjust the dimension of the features through a linear transformation layer and a transpose operation, and output the distance and direction matrices respectively; For the residue direction matrix, with the Pair feature as the input, construct a structure symmetry state deduction network guided by dilated convolution for prediction, including the following steps: S21: Add the Pair feature to its transpose matrix and take the average to obtain a symmetric matrix as the input; S22: Pass through multiple convolutional modules, including a convolutional layer, an instance normalization layer, and an ELU activation function layer, add the features output by multiple convolutional layers to the original input and activate through ELU to implement residual connection; S23. Adjust the dimension through a convolutional layer and a linear transformation layer, and output a dihedral angle with a shape of 1, nres, nres, 37 after the transpose operation, where nres is the number of amino acid residues in the antibody sequence. Orientation matrix.

2. The method for predicting the dual-branch structural features oriented to antibodies according to claim 1, wherein: In S13 , , , projected onto 37, projected onto 18.

3. The method for predicting the dual-branch structural features oriented to antibodies according to claim 1, wherein: In S22, the convolutional layer sequentially uses convolutional kernels with increasing sizes and dilation rates for dilated convolution to expand the receptive field.

4. A method for predicting the dual-branch structural characteristics oriented to antibodies according to claim 1, characterized in that: In step 3 of model training, the RoseTTAFold2 network is used to extract features from the antibody sequence.

5. A method for predicting the dual-branched structural features of an antibody according to claim 4, wherein: In step 3 of model training, the activation functions used are the exponential linear unit ELU and the rectified linear unit ReLU, the dropout is set to 0.1 to avoid overfitting, the optimizer used is Adam, trained for 10 epochs, the learning rate is set to 0.001, and the learning rate scheduler MultiStepLR is used to reduce the learning rate by 0.5 times in the 2nd, 3rd, 4th, 6th, and 8th epochs.

Images