Sequence-based antigen antibody affinity prediction method

Abstract

The application discloses a sequence-based antigen antibody affinity prediction method, comprising the following steps: S1. obtaining the amino acid sequences of the antigen and the antibody and the affinity index from the TDC database; there are 431 antibodies and 401 antigens in the 493 data pairs; the obtained sequence is the full-length sequence of the antigen and the antibody, for example, KVFGRAKFESNFATNRNTDGGILQIN, and each capital letter represents an amino acid residue; S2. each piece of information also corresponds to a kd value, and the unit is mol / L; each kd value is converted into a -pkd as a label; S3. the data set is evenly divided into five parts, and the same antigen and antibody are ensured to be in the same part for five-fold cross-validation training; only the partial sequence of the CDR region of the antibody sequence is obtained, and only the partial sequence of the epitope obtained through prediction of the antigen sequence, so that the key binding information is obtained during neural network training, and the training amount is reduced; meanwhile, the embedding representation obtained by combining the natural language pre-training model of the protein sequence is used to improve the model accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field related to biological medicine, and particularly relates to a sequence-based antigen antibody affinity prediction method. BACKGROUND

[0002] Antibodies are a special protein produced by the immune system, which can bind to viruses and other substances invading the human body, i.e. antigens, to inactivate them. Compared with small molecule drugs, antibody drugs can act on more target proteins, and can realize high binding affinity and specific reduction of toxicity through protein engineering technology. Now there are many antibody drugs on the market. The binding force between antibody and antigen epitope or antigen determinant is called antibody affinity, which is essentially a non-covalent force, including the attraction between amino acids, hydrogen bonds, hydrophobic force, etc. Antibody affinity reflects the ability of a determinant of an antibody molecule and an antigen molecule to react. Improving the affinity of antibodies and antigens is an important step in the later stage of antibody research and development. Since traditional antibody design and optimization requires wet experiments, it is time-consuming and laborious, and the prediction of antibody-antigen affinity mostly requires structure, which requires higher requirements, so most methods for predicting antigen-antibody affinity rely on modeling based on three-dimensional structure information of proteins. If there is no complex structure, the antigen and antibody need to be docked first, which reduces the prediction accuracy. Secondly, sometimes the structure of antigen and antibody is not clear, only the sequence information is available. SUMMARY

[0003] The purpose of the present application is to provide a sequence-based antigen antibody affinity prediction method to solve the problems of reducing prediction accuracy and sometimes the structure of antigen and antibody is not clear, only the sequence information is available as mentioned in the background.

[0004] To achieve the above purpose, the present application provides the following technical scheme: a sequence-based antigen antibody affinity prediction method, comprising the following steps:

[0005] S1. Obtain the amino acid sequences of antigen and antibody and the affinity index from the TDC database; there are 493 data pairs, of which 431 antibodies and 401 antigens are actually available; the obtained sequence is the full-length sequence of antigen and antibody, for example, KVFGRAKFESNFATNRNTDGGILQIN, and each capital letter represents an amino acid residue;

[0006] S2. Each piece of information also corresponds to a kd value, with a unit of mol / L; each kd value is converted into a -kd as a label;

[0007] S3. Divide the data set into five equal parts, and ensure that the same antigen and antibody are in the same part for five-fold cross-validation training;

[0008] S4. Calculate the average kd value of each antigen and antibody in each part, and use the average kd value as the final kd value of the antigen and antibody in the part;

[0008] S4. The neural network is built and trained using the Keras library; the dimension of the GRU output space is set to 512, the optimizer is sgd, the training is performed for 100 times, the learning rate is 0.01, and the batch size is 512;

[0009] S5. The predicted output of the model -pkd value is converted into affinity value, unit kcal / mol;

[0010] S6. A separate test set is set up, with different antigen-antibody sequences and affinity data of 20 in the training set;

[0011] S7. Five-fold cross-validation results and independent test set validation results (Table 1), calculate the correlation of true and predicted affinity.

[0012] Preferably, the flexibility of the antigen and antibody binding is determined by the antibody complementarity determining region (CDR), which consists of six fragments, three fragments for the heavy chain and three fragments for the light chain of the antibody. First, the AI-based prediction method abnumber program is used to predict the CDR region of the antibody, which can accurately predict the CDR region of the antibody. Only by inputting the heavy chain sequence and light chain sequence of the antibody, the CDR region sequence can be extracted, and the six CDR regions of the antibody are combined together while ignoring the sequence part of other antibodies which has little contribution to the antigen binding.

[0013] Preferably, the part of the antigen and antibody binding is called antigen epitope, although the structure of the antigen is very large, but the amino acid residues of the antigen epitope play a decisive role in the binding with the antibody; in order to obtain the sequence of these amino acid residues, Bepipred3.0-predictor is used, which is a B cell epitope prediction tool based on ESM-2 encoding according to protein sequence. Only by inputting the sequence of the antigen, the tool will output the predicted epitope sequence.

[0014] Preferably, the fragments linked by the six CDR regions and the antigen epitope region are all composed of amino acid sequences, which are first encoded into vectors that can be recognized by the neural network; the pre-trained model ESM-2 (Evolutionary Scale Modeling) and TAPE (Tasks Assessing Protein Embeddings) of facebook protein language are used for feature embedding (Embedding).

[0015] Preferably, the TAPE model uses a bidirectional encoder representation from a transformers model and is trained on 31 million protein sequences from Pfam and validated through six downstream tasks (remote homology detection, secondary structure prediction, contact prediction, and three protein engineering tasks).

[0016] Preferably, the ESM model also uses a transformers model but is trained on a total of 250 million protein sequences and tested on three benchmarks: fold-level remote homology, secondary structure prediction, and top-L remote contact precision.

[0017] Preferably, in the neural network, the embedding features are input to the RCNN network used and connected to the output head, which is composed of convolutional layers and GRUs in addition;

[0018] In the structure of each GRU unit, h t represents the output at time t, xt represents the input at time t, ht-1 represents the output of the last time, r represents the reset gate, and "update" represents the update gate, where the gate represents matrix operation,

[0019]

[0020]

[0021] where "sigmoid" represents the activation function, con represents the concatenation operation on the input matrix, and W represents the weight factor matrix of the reset gate Wr and the update gate Wz; h / and ht can be represented as follows

[0022]

[0023]

[0024] where the dot is a matrix multiplication operation, which can be represented as a function of the output ht-1 of the last step; 1-z also indicates that ht at the current step t contains the information of the last step.

[0025] Compared with the prior art, the application provides a sequence-based antigen antibody affinity prediction method, which has the following beneficial effects:

[0026] The application only obtains part of the sequence of the CDR region for the antibody sequence, and only obtains the epitope part sequence after prediction for the antigen sequence, so that the key binding information is obtained during neural network training, and the training amount is reduced; at the same time, the embedding representation obtained by combining the natural language pre-training model of the protein sequence improves the accuracy of the model. BRIEF DESCRIPTION OF DRAWINGS

[0027] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention, but do not constitute a limitation thereof. In the drawings:

[0028] Figure 1 This is a schematic diagram of a GRU network;

[0029] Figure 2 A schematic diagram of an antigen-antibody sequence affinity prediction model;

[0030] Figure 3 Table 1 shows the correlation coefficients of the model on the training, validation, and independent test sets. Detailed Implementation

[0031] 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.

[0032] Please see Figures 1-3 This invention provides a technical solution: a sequence-based method for predicting antigen-antibody affinity, characterized by comprising the following steps:

[0033] S1. The amino acid sequences and affinity indices of antigens and antibodies were obtained from the TDC database; a total of 493 data pairs were obtained, including 431 antibodies and 401 antigens; the obtained sequences were the full-length sequences of antigens and antibodies, such as KVFGRAKFESNFATNRNTDGGILQIN, where each capital letter represents an amino acid residue.

[0034] S2. Each piece of information also corresponds to a kd value, in mol / L; each kd value is converted into -pkd as a tag;

[0035] S3. Divide the dataset into five equal parts, and ensure that the same antigens and antibodies are used for five-fold cross-validation training in the same part;

[0036] S4. The neural network was built and trained using the Keras library; the GRU output space dimension was set to 512, the optimizer was sgd, training was performed 100 times, the learning rate was 0.01, and the batch size was 512.

[0037] S5. Convert the -pkd value of the model prediction output into an affinity value, in kcal / mol;

[0038] S6. A separate test set was set up, with different antigen-antibody sequences and affinity data from the training set, 20 pieces;

[0039] S7. Five-fold cross-validation results and independent test set validation results (Table 1), calculate the correlation between true and predicted affinity, the average correlation coefficient on the independent test set is 0.54, compared with the correlation coefficient 0.4 obtained by the existing method CSM-AB, it has certain improvement.

[0040] As Figure 2 shown, the flexibility of antigen and antibody binding is determined by the antibody complementarity determining region (CDR), which consists of six fragments, three fragments for the heavy chain and light chain of the antibody respectively, first, the AI-based prediction method abnumber program is used to predict the CDR region of the antibody, which can accurately predict the CDR region of the antibody, only by inputting the heavy chain sequence and light chain sequence of the antibody, the CDR region sequence will be extracted, and the 6 CDR regions of the antibody are combined together and the sequence part of other antibodies which has small contribution to the antigen binding is ignored; The part of antigen and antibody binding is called antigen epitope, although the structure of antigen is very large, but the amino acid residues of antigen epitope play a decisive role in the binding with antibody; In order to obtain the sequence of these amino acid residues, Bepipred3.0-predictor is used, which is a B-cell epitope prediction tool based on ESM-2 encoding according to protein sequence, only by inputting the sequence of antigen, the tool will output the predicted epitope sequence, the fragments linked by 6 CDR regions and the antigen epitope region are all composed of amino acid sequences, first, the amino acid sequence is encoded into a vector that can be recognized by neural network; Facebook's protein language pre-training model ESM-2 (Evolutionary Scale Modeling) and TAPE (Tasks Assessing Protein Embeddings) are used for feature embedding (Embedding), TAPE model uses bidirectional encoder representation from transformers model, and is trained on 31 million protein sequences from Pfam, and is verified by six downstream tasks (remote homology detection, secondary structure prediction, contact prediction and three protein engineering tasks), ESM model also uses transformers model, but is trained on a total of 250 million protein sequences, and is tested in three benchmark tests: fold-level remote homology, secondary structure prediction and top-L remote contact precision.

[0041] As Figure 1 shown, in the neural network, the embedded features are input to the RCNN network used and connected to the output head, the network is composed of convolutional layers and GRU addition;

[0042] In the structure of each GRU unit, h t xt represents the input at time t, ht-1 represents the output of the last time, r represents the reset gate, "update" represents the update gate, and the gate here represents the matrix operation,

[0043]

[0044]

[0045] where "sigmoid" represents the activation function, con represents the concatenation operation on the input matrix, and W represents the weight factor matrix of the reset gate Wr and the update gate Wz; h / and ht can be represented as follows

[0046]

[0047]

[0048] where the dot is the matrix multiplication operation, which can be represented as a function of the output of the last step ht-1; 1-z also indicates that ht contains the information of the last step in the current step t.

[0049] Usage: First, initialize the module, activate the python environment, import various dependent file toolkits required for software running, switch to the folder where the program is located, run the python file to input the antibody sequence and antigen sequence to be predicted, and then the software predicts and outputs the prediction value.

[0050] Although embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A sequence-based method for predicting antigen-antibody affinity, characterized in that, Includes the following steps: S1. The amino acid sequences and affinity indices of antigens and antibodies were obtained from the TDC database; a total of 493 data pairs were obtained, including 431 antibodies and 401 antigens; the obtained sequences were the full-length sequences of antigens and antibodies, with each capital letter representing an amino acid residue. S2. Each piece of information also corresponds to a kd value, in mol / L; each kd value is converted into -pkd as a tag; S3. Divide the dataset into five equal parts, and ensure that the same antigens and antibodies are used for five-fold cross-validation training in the same part; S4. The neural network was built and trained using the Keras library; the GRU output space dimension was set to 512, the optimizer was sgd, training was performed 100 times, the learning rate was 0.01, and the batch size was 512. S5. Convert the -pkd value of the model prediction output into an affinity value, in kcal / mol; S6. An independent test set was set up, containing 20 antigen and antibody sequences and affinity data that were different from the training set. S7. Five-fold cross-validation results and independent test set validation results (Table 1) are used to calculate the correlation between true and predicted affinity.

2. The sequence-based antigen-antibody affinity prediction method according to claim 1, characterized in that: The flexibility of antigen-antibody binding is determined by the antibody complementarity-determining region (CDR), which consists of six segments: three segments from the heavy chain and three from the light chain. First, the CDR region of the antibody is predicted using the AI-based prediction method abnumber program. To predict the CDR region, only the heavy chain and light chain sequences of the antibody need to be input, and the CDR region sequence will be extracted. The CDR regions of the six antibodies are combined together, while the sequence parts of other antibodies that contribute little to antigen binding are ignored.

3. The sequence-based antigen-antibody affinity prediction method according to claim 2, characterized in that: The part of the antigen that binds to the antibody is called the antigenic epitope. Although the antigen has a large structure, the amino acid residues of the antigenic epitope play a decisive role in binding to the antibody. In order to obtain the sequence of these amino acid residues, Bepipred3.0-predictor is used. This is a B-cell epitope prediction tool based on ESM-2 encoding of protein sequences. Just input the sequence of the antigen, and the tool will output the predicted epitope sequence.

4. The sequence-based antigen-antibody affinity prediction method according to claim 3, characterized in that: The fragments linked by the six CDR regions and the antigenic epitope regions are all composed of amino acid sequences. First, the amino acid sequences need to be encoded into neural network recognition vectors. Facebook's protein language pre-trained model ESM-2 and TAPE (TasksAssessingProteinEmbeddings) were used for feature embedding.

5. The sequence-based antigen-antibody affinity prediction method according to claim 4, characterized in that: The TAPE (TasksAssessingProteinEmbeddings) uses a bidirectional encoder representation from a transformers model and is trained on 31 million protein sequences from Pfam. It is validated through six downstream tasks: long-range homology detection, secondary structure prediction, contact prediction, and three protein engineering tasks.

6. The sequence-based antigen-antibody affinity prediction method according to claim 5, characterized in that: The ESM-2 model also uses transformers, but was trained on a total of 250 million protein sequences and tested on three benchmarks: long-range homology at the fold level, secondary structure prediction, and top-L long-range contact accuracy.

7. The sequence-based antigen-antibody affinity prediction method according to claim 1, characterized in that: In the neural network, the embedded features are input into the RCNN network and connected to the output head. The RCNN network consists of convolutional layers and GRUs added together. In the structure of each GRU unit, h t The output at time t, x t h represents the input at time t. t-1 This indicates the previous output, 'r' represents the reset gate, and "update" represents the update gate, where gates represent matrix operations. Where "sigmoid" represents the activation function, con represents the cascade operation on the input matrix, and W represents the reset gate. Update Gate The weight factor matrix; and h t It is expressed as follows The dot represents matrix multiplication, denoted as the output h from the previous step. t-1 The function; 1-z also represents h in the current step t. t It contains information from the previous step.

Images