A method for predicting antibody-like binding to biomacromolecule modification based on protein domain

Through the method of predicting the binding to biological macromolecule modification by protein domain prediction, high-throughput sequencing and convolutional neural networks are used to screen out antibodies that bind to tyrosine phosphorylation modification, which solves the long cycle and high cost of the existing technology. and the problem of missing low copy number antibodies has been achieved, which has achieved rapid and efficient antibody-like screening and discovery.

CN117594114BActive Publication Date: 2024-12-03UNIVERSITY OF HEALTH & REHABILITATION SCIENCES +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311419536.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-10-30
Publication Date
2024-12-03
Estimated Expiration
2043-10-30

AI Technical Summary

Technical Problem

When the prior art discovers antibodies that bind to biological macromolecules modification, the method cycle is long and costly, and it is easy to miss antibodies with low copy number, making it difficult to efficiently screen out mutants with high affinity.

Method used

Through the method of predicting protein domain binding to biological macromolecule modification, high-throughput sequencing and convolutional neural networks are used to construct antibody-like prediction models, and antibodies that bind to tyrosine phosphorylation are screened to achieve rapid screening and high precision Discover.

Benefits of technology

It has achieved rapid exploration of the characteristics of antibody-like sequences, and screened out high-affinity antibodies with low cost and efficient use, filling the gap in low copy number antibodies that are difficult to detect in the existing technology, and improving discovery efficiency and accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117594114B_ABST
    Figure CN117594114B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for generating antibody-like proteins based on protein domain prediction combined with biomacromolecule modification. The method of the present invention includes the construction, panning and high-throughput sequencing of a protein domain mutation library, the processing of high-throughput sequencing data and the construction of a training data set for an antibody-like prediction model, the construction and training of an antibody-like prediction model based on a convolutional neural network, and the use of the model to predict and screen candidate antibody-like proteins to explore characteristic sequences. The method of the present invention can be used to explore the characteristics of antibody-like sequences and to rapidly screen antibody-like proteins, which is beneficial for finding high-affinity antibody-like proteins that are difficult to find using phage display or biopanning techniques, and has the advantages of low cost and high precision.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics technology and more specifically relates to a method for predicting antibody-like substances that bind to biomacromolecule modifications based on protein domains. Background Art

[0002] Biomacromolecule modification refers to the covalent bonding of small molecule groups to the bases of DNA, RNA or the amino acid side chains of proteins, thereby regulating the function of biomacromolecules. The development of antibodies that bind to biomacromolecule modification helps to accurately detect pathological conditions caused by changes in biomacromolecule modification sites, providing assistance for the diagnosis and treatment of diseases.

[0003] Quasi-antibodies refer to the mutation of proteins that can bind to biomacromolecule modifications (such as SH2 domains that bind to tyrosine phosphorylation modifications or Chromo domains that bind to lysine methylation modifications). The generated mutants have a much higher affinity for biomacromolecule modifications than the wild type, similar to the binding level of antigen-antibody (nM or pM level). In short, quasi-antibodies are proteins that can bind to antigens with high affinity, similar to the functions of antibodies. Since the groups modified by biomacromolecules are small, it is very difficult to develop antibodies that can efficiently bind and enrich the modified groups. However, for the domains of proteins known to bind to biomacromolecule modifications, through protein engineering, it may be possible to develop quasi-antibodies that bind to the modified groups with high affinity. For example, a mutant library of the target protein is constructed through phage display technology, and the library is screened with the modified groups to discover quasi-antibodies.

[0004] At present, there is a method that uses directed evolution as a research method and SH2 domain as a research object to explore whether the tyrosine phosphorylation modification (phosphorylated tyrosine, pY) binding pocket in the SH2 protein structure can be improved by large-scale mutation (SH2 superbinder) (Kaneko, T., et al. (2012). "Superbinder SH2 domains act as antagonists of cell signaling." Science signaling 5 (243): ra68.). Similarly, the directed evolution method is also used to discover pY antibodies (Li, S., et al. (2021). "Revisiting the phosphotyrosine binding pocket of Fyn SH2 domainled to the identification of novel SH2 superbinders."Protein Science 30(3):558-570). However, such methods require the use of phage display and biopanning technology to pan the phage displayed mutant library, which has a long experimental cycle and high cost. Although this method can discover antibodies with a higher copy number, it is easy to miss antibodies with a lower copy number. Therefore, it is necessary to provide a method for efficiently discovering antibodies. Summary of the invention

[0005] In view of the above technical problems, the present invention provides a method for predicting antibody-like substances binding to biomacromolecule modifications based on protein domains, and also provides antibody-like substances binding to tyrosine phosphorylation-modified peptides obtained based on the method.

[0006] The first object of the present invention is to provide a method for predicting antibody-like substances that bind to biomacromolecule modifications based on protein domains.

[0007] The second object of the present invention is to provide a system for predicting antibody-like proteins that bind to biomacromolecule modifications based on protein domains.

[0008] The third object of the present invention is to provide an antibody-like substance that binds to a peptide containing tyrosine phosphorylation modification and its application.

[0009] The above-mentioned purpose of the present invention is achieved through the following technical solutions:

[0010] The present invention provides a method for predicting antibody-like substances that bind to biomacromolecule modifications based on protein domains, comprising the following steps:

[0011] S1. Based on the protein domain, analyze the amino acid sites that may bind to biomacromolecule modification and mutate them, obtain the mutant library and then perform panning;

[0012] S2. Perform high-throughput sequencing on the mutant library and the library obtained by panning, process the sequencing data, obtain mutant sequences and construct an antibody-like prediction model data set; the antibody-like prediction model data set includes an antibody-like binary classification model and a regression model data set;

[0013] S3. constructing an antibody-like binary classification prediction model and a regression prediction model based on a convolutional neural network, and using the constructed binary classification model and regression model data sets to train the constructed binary classification prediction model and regression prediction model respectively;

[0014] S4. Use the trained binary classification prediction model and regression prediction model to score the mutant sequences obtained in step S2, retain the mutant sequences with higher scores than the wild type, count the enriched amino acids at each mutation position of the mutant sequence, arrange and combine the amino acids according to the order of the mutation positions, and generate a sequence to be predicted;

[0015] S5. Calculate the conservation of the amino acid at each mutation position in the sequence to be predicted, and group the sequences to be predicted according to the amino acid at the least conserved mutation position; recalculate the conservation of the amino acid at each mutation position of the sequence to be predicted in each group, determine the most conserved amino acid at each position in each group, combine the most conserved amino acids in sequence, and the formed sequence is the predicted antibody-like protein.

[0016] As one embodiment, when predicting an antibody-like antibody that can bind to a peptide containing a tyrosine phosphorylation modification based on mutation of the Fyn SH2 domain, the mutated 8 amino acid sites correspond to Lys19, Glu41, Thr42, Thr43, Ala46, Ser48, Lys63 and Lys66 of the parent Fyn SH2 domain (shown in SEQID NO.1), respectively.

[0017] Specifically, a mutant library obtained by mutation of the SH2 domain was panned using peptides containing tyrosine phosphorylation modifications.

[0018] Specifically, in the method described in the present invention, the processing of the sequencing data described in step S2 includes preliminary processing and denoising processing; wherein the preliminary processing is: splicing the DNA sequence obtained by converting the high-throughput sequencing data and performing quality control, setting the quality control value of a single base to be not less than 20, retaining DNA sequences in which the base ratio that meets this quality control value is not less than 90%, and translating the retained DNA sequences into amino acid sequences; the denoising processing is: deleting the amino acid sequence having a copy number less than 3 and not present in all rounds of panning; after processing, the mutant sequence is obtained.

[0019] Specifically, when translating a DNA sequence into an amino acid sequence, the stop codon is also translated; among them, the codon TAG is translated into q; the codon TAA is translated into x; and the codon TGA is translated into w.

[0020] Specifically, the amino acids at each mutation site on the mutant sequence in the library before and after panning are extracted and sequentially connected into peptide segments for the construction of an antibody-like prediction model data set.

[0021] Specifically, the binary classification prediction model data set in step S2 includes two columns of data, an amino acid sequence composed of amino acids at the mutation positions of the mutant sequence and a label column; the value of the label column is defined as 0 or 1; the definition method of the label value of the amino acid sequence is:

[0022] (1) Calculate the proportion of the amino acid sequence copy number in the total copy number of the sequence in each panning round;

[0023] (2) Calculate the difference in the proportion of the amino acid sequence in the two rounds of data before and after the panning; define the amino acid sequence as follows based on the difference in proportion:

[0024] If the proportion of the amino acid sequence after panning is greater than the proportion before panning, the label value of the amino acid sequence is defined as 1, which is a positive sample; otherwise, it is 0, which is a negative sample. When the number of positive and negative samples is unequal, select from the maximum and minimum values ​​of the proportion value to the middle to make the number of positive and negative samples equal.

[0025] The regression model data set includes two columns of data, an amino acid sequence composed of amino acids at the mutation positions of the mutant sequence and a label column; the value of the label column is defined as follows:

[0026] (1) Calculate the proportion of the amino acid sequence copy number in the total copy number of each round of sequences;

[0027] (2) Calculate the ratio of the proportion after panning to the proportion before panning, and perform a log10 operation on the ratio; when the proportion is 0, set the copy number of the sequence with a copy number of 0 to 10, and recalculate the proportion;

[0028] (3) Taking 0 as the boundary, the label values ​​are divided into two categories according to the log10 operation results. The number of sequences in the smaller category is counted as n, and the number of sequences in the larger category is counted as m. The sequences in the smaller category are included in the regression prediction model data set; the same number of sequences are selected from the larger category and also included in the regression prediction model data set; the selection method is: arrange the m sequences according to the size of the label value, divide them into n parts, randomly select a sequence from each part, and finally obtain n sequences, which are included in the regression prediction model data set.

[0029] Specifically, the method for constructing the binary classification prediction model and the regression prediction model in step S3 is as follows:

[0030] (1) Determine feature encoding: Input the feature vector generated after feature encoding into the input layer, use the algorithm to test the commonly used protein encoding methods, and evaluate the performance of different feature encodings through cross-validation. Select the encoding with the best performance as the feature encoding of the data set;

[0031] (2) constructing input layer, convolution layer, fully connected layer and output layer models: the difference between the models is only the number of convolution layers, and the number of convolution layers of the models is 2 to 6;

[0032] The convolution kernel size of the first convolution layer is fixed to 1, the convolution kernel size of the remaining convolution layers is fixed to 3, the number of convolution kernels of all convolution layers is fixed to 64, and a dropout layer is added after all convolution layers, and the parameter is fixed to 0.5.

[0033] The number of fully connected sublayers in the fully connected layer is fixed to 3, the number of units in the last fully connected sublayer is set to 1, the number of units in the remaining fully connected sublayers is fixed to 32 and a dropout layer is added at the end with a fixed parameter of 0.5, and there is a global average pooling layer before the last fully connected sublayer;

[0034] The output layer includes one neuron and uses sigmoid as the activation function; the patience of the early stopping strategy is fixed to 100, the batch_size value is fixed to 256, and the epoch parameter is fixed to 1000.

[0035] Specifically, when constructing the input layer, convolutional layer, fully connected layer, and output layer models, the performance of different models is evaluated through cross-validation, and the model framework with the best performance is selected for further optimization.

[0036] Specifically, the optimization includes:

[0037] (1) Optimization of the number of convolution kernels: The convolution kernel size of each convolution layer is set to 64 or 128, where the convolution kernel size of the next convolution layer is not less than that of the previous layer. Different models are evaluated through cross-validation to determine the number of convolution kernels for each convolution layer.

[0038] (2) Optimization of convolution kernel size: The convolution kernel size of all convolution layers except the first layer is set to 2 to 10, where the convolution kernel size of the next layer is not less than that of the previous layer. Different models are evaluated by cross-validation to determine the convolution kernel size of each convolution layer.

[0039] (3) Optimization of the maximum pooling layer of the convolutional layer: Add a maximum pooling layer after each convolutional layer from the back to the front. Use cross-validation to evaluate different models and determine the number of maximum pooling layers to add.

[0040] (4) Optimization of the number of fully connected sub-layers: The number of fully connected sub-layers was set to 3 to 7 respectively, and different models were evaluated by cross-validation to determine the number of fully connected sub-layers.

[0041] (5) Optimization of the number of fully connected sub-layer units: The number of fully connected sub-layer units except the last layer is set to 4, 8, 32, 64, 128 or 256, where the number of fully connected sub-layer units in the previous layer is not less than the number of fully connected sub-layer units in the next layer. Different models are evaluated by cross-validation to determine the number of units in each fully connected sub-layer.

[0042] (6) Optimization of the maximum pooling layer of the fully connected layer: Except for the last two fully connected sub-layers, a maximum pooling layer is added after each sub-layer. The cross-validation method is used to evaluate whether a maximum pooling layer needs to be added.

[0043] (7) Dropout layer parameter optimization: Set each dropout layer parameter to 0.2 to 0.8. Use cross-validation to evaluate different models and determine the parameters of each dropout layer.

[0044] (8) Early stopping strategy: Set the patience value to 30, 50, 70, 90, 110, 130 or 150. Evaluate different models through cross-validation to determine the optimal patience value.

[0045] (9) Batch_size optimization: Set the batch_size value to 32, 64, 128, 256 or 512. Evaluate different models through cross-validation to determine the optimal batch_size value.

[0046] Specifically, the present invention uses One-Hot encoding to convert the amino acid sequence in the data set into a digital matrix that can be recognized by the CNN model; in One-Hot encoding, 20 amino acids and are encoded as a 20-dimensional binary vector, and in the vector corresponding to the amino acid, the elements related to the amino acid are marked as 1, and the other elements are marked as 0.

[0047] Specifically, in step S4, when retaining mutant sequences with scores higher than the wild type, the threshold of the binary prediction model can be set to 0.5, and sequences with binary prediction model scores higher than 0.5 and regression prediction model scores greater than the wild type are retained.

[0048] Specifically, in step S4, when determining the enriched amino acids, the proportion of amino acids at each mutation position is counted, and the amino acids at each position with a proportion greater than 0.05 are defined as enriched amino acids.

[0049] At the same time, a Two-Sample-Logo diagram can also be drawn, and the enriched amino acids obtained from the diagram can also be included.

[0050] Specifically, the rule for generating the sequence to be predicted in step S4 is as follows: one of the mutation positions is allowed to be a common amino acid, and the remaining positions must be enriched amino acids. For example, there are 8 positions for mutation based on the SH2 domain in the present invention, of which one mutation position is a common amino acid (20 species), and the remaining positions are enriched amino acids, thereby generating 5,664,001 sequences to be predicted.

[0051] Specifically, if the number of generated sequences to be predicted is large, the trained binary classification prediction model and regression prediction model can be continued to be used to score the generated sequences to be predicted, and the sequences to be predicted whose scores of the binary classification prediction model and the regression prediction model are higher than those of the wild type are retained.

[0052] Specifically, if the number of sequences to be predicted with a binary prediction model and a regression prediction model score higher than the wild type is still large, they can be further filtered. Since the lower the score of the binary prediction model and the regression prediction model score, the lower the affinity, the sequences to be predicted with a binary prediction model score>0.9 and a regression prediction model score>1 can be selected and retained, which are called candidate antibodies.

[0053] The present invention also provides a system for predicting antibody-like compounds combined with biomacromolecule modifications based on protein domains, comprising a high-throughput sequencing data processing module, an antibody-like prediction model evaluation module, and a prediction sequence output module;

[0054] The high-throughput sequencing data processing module is used to process high-throughput sequencing data and construct an antibody-like binary classification model and a regression model data set, and submit the data set to the antibody-like prediction model evaluation module;

[0055] The antibody-like prediction model evaluation module includes a binary classification prediction model and a regression prediction model of the antibody-like, firstly trains the binary classification prediction model and the regression prediction model using the input binary classification model and regression model data sets, then inputs the sequence to be evaluated into the binary classification prediction model and the regression prediction model, and outputs the binary classification prediction model score and the regression prediction model score;

[0056] The predicted sequence output module screens the sequences according to the input binary prediction model scores and the regression prediction model scores, retains the sequences with scores higher than the wild type, counts the enriched amino acids at each mutation position of the sequence, arranges and combines the amino acids according to the order of the mutation positions, generates the sequence to be predicted and calculates the conservation of the amino acids at each mutation position in the sequence to be predicted, groups the sequences to be predicted according to the amino acids at the least conserved mutation position; recalculates the conservation of the amino acids at each mutation position of the sequence to be predicted in each group, determines the most conserved amino acid at each position in each group, combines the most conserved amino acids in sequence, forms and outputs the predicted antibody-like sequence.

[0057] The present invention also provides an antibody-like compound that binds to a peptide containing tyrosine phosphorylation modification, wherein the antibody-like compound is predicted by the method of the present invention; the amino acid sequence of the antibody-like compound is shown in any one of SEQ ID NOs. 2 to 11.

[0058] Specifically, the antibodies described in the present invention are M_4V2, M_4V4, M_4R1, M_4R2, M_4R4, M_4F1, M_4F2, M_4N, M_4V1 and M_4F2, and their amino acid sequences are shown in SEQ ID NOs. 2 to 11 respectively.

[0059] The present invention claims to protect the use of the antibody-like substance that binds to the peptide containing tyrosine phosphorylation modification in the preparation of a product for detecting the peptide containing tyrosine phosphorylation modification.

[0060] Specifically, the detection of peptides containing tyrosine phosphorylation modifications includes qualitative detection and quantitative detection, that is, it can also be used to detect the concentration of peptides containing tyrosine phosphorylation modifications.

[0061] In view of the fact that the antibody-like substance that binds to the peptide containing tyrosine phosphorylation modification described in the present invention can be used to detect the peptide containing tyrosine phosphorylation modification, and the corresponding DNA sequence, i.e., the DNA sequence encoding the antibody-like substance described in the present invention, can be obtained based on the amino acid sequence provided by the present invention, so as to obtain the antibody-like substance described in the present invention through recombinant expression, etc. Therefore, the present invention also requests protection for the DNA sequence encoding the antibody-like substance described in the present invention.

[0062] The present invention also claims to protect the vector containing the DNA sequence encoding the antibody of the present invention.

[0063] The present invention has the following beneficial effects:

[0064] The present invention provides a method for generating antibody-like proteins based on protein domain prediction combined with biomacromolecule modification, including the construction, panning and high-throughput sequencing of a protein domain mutation library, the processing of high-throughput sequencing data and the construction of a training data set for an antibody-like prediction model, the construction and training of an antibody-like prediction model based on a convolutional neural network, and the use of the prediction model to screen candidate antibody-like proteins to explore characteristic sequences. The method of the present invention can be used to explore the characteristics of antibody-like sequences and to rapidly screen antibody-like proteins, which is beneficial for finding high-affinity antibody-like proteins that are difficult to find using phage display or biopanning techniques, has the advantages of low cost and high precision, and is of great significance for the prediction and screening of antibody-like proteins. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 The present invention is a flow chart for obtaining antibody-like substances modified and bound to biological macromolecules using the method of the present invention.

[0066] Figure 2 The positions of the eight amino acid sites on the Fyn SH2 domain that bind to pTyr.

[0067] Figure 3 The change of OD450 signal of SH2 mutant library during bio-panning.

[0068] Figure 4 It is the TSL graph of the positive and negative sample data set of the binary classification prediction model.

[0069] Figure 5 This is the network architecture diagram of the antibody prediction model.

[0070] Figure 6 A scatter plot showing the relationship between the predicted scores of the regression model for an independent test set and the true values.

[0071] Figure 7 This is the Sequence Logo diagram of potential pY class antibodies.

[0072] Figure 8 It is a Sequence Logo diagram of the sequence of the amino acid fixed at the P4 position; A to I in the figure are Sequence Logo diagrams of the sequence of the amino acids fixed at the 4th position, namely R (A), N (B), F (C), D (D), T (E), E (F), V (G), W (H) and I (I), respectively.

[0073] Fig. 9 To predict the ELISA experimental results of antibody-like and phosphorylated peptide pYEEI.

[0074] Fig.10 The ELISA results of the antibody and phosphorylated peptide pYEEI for comparison.

[0075] Fig.11 These are the affinity test results of predicted antibodies and phosphorylated peptide pYEEI; A to L in the figure are the affinity test results of predicted antibodies M_4V2, M_4V3 (Trm), M_4V4, M_4R1, M_4R2, M_4R3 (V13), M_4R4, M_4F1, M_4S1, M_4N, M_4N1, M_4N2 and phosphorylated peptide pYEEI, respectively.

[0076] Fig.12 These are the affinity test results of the predicted antibodies and phosphorylated peptide GGpYGG; A to L in the figure are the affinity test results of the predicted antibodies M_4V2, M_4V3 (Trm), M_4V4, M_4R1, M_4R2, M_4R3 (V13), M_4R4, M_4F1, M_4S1, M_4N, M_4N1, M_4N2 and phosphorylated peptide GGpYGG, respectively. DETAILED DESCRIPTION

[0077] The present invention is further described below in conjunction with the accompanying drawings and specific examples, but the examples do not limit the present invention in any form. Unless otherwise specified, the reagents, methods and equipment used in the present invention are conventional reagents, methods and equipment in the art.

[0078] Unless otherwise specified, the reagents and materials used in the following examples are commercially available.

[0079] The present invention uses the prediction of pY class antibodies as an example to demonstrate the reliability of the method of predicting class antibodies that bind to biomacromolecule modifications based on protein domains described in the present invention. The flowchart of obtaining class antibodies that bind to biomacromolecule modifications using the method described in the present invention is shown in FIG. Figure 1 As shown, the process includes the following five steps: data collection and processing, construction of an antibody-like prediction model, prediction of antibody-like antibodies, exploration of antibody-like sequence features, and experimental verification.

[0080] Example 1 Construction of Class 1 Antibody Prediction Model Training Dataset

[0081] The mutation library constructed in the reference document (Li S, et al. Revisiting the phosphotyrosine binding pocket of Fyn SH2 domain led to the identification of novel SH2superbinders. Protein Sci, 2021, 30 (3): 558-70.) of the present invention uses the high-throughput sequencing data of the mutation library and the library obtained after panning as training data to train the pY class antibody prediction model. Specifically, the mutation library is obtained by random mutation of 8 amino acid sites on the Fyn SH2 domain. The positions of the 8 amino acid sites on the Fyn SH2 domain that bind to pTyr are as follows Figure 2 On this basis, a bio-panning experiment was performed using phage display technology, and a synthetic peptide containing pTyr was used for affinity testing. The synthetic peptide containing pTyr used to experimentally verify the affinity of the Fyn SH2 mutant is shown in Table 1. The N-terminus of the peptide is biotinylated and the C-terminus is amidated.

[0082] Table 1. Synthetic peptides containing pTyr used to experimentally verify the affinity of Fyn SH2 mutants

[0083]

[0084]

[0085] Note: “a”: 6-aminoacetic acid; “amide”: amidated C-terminus (-CONH2); “pY”: phosphorylated tyrosine

[0086] The created mutant library and the library after panning were subjected to high-throughput sequencing using the Illumina second-generation high-throughput sequencing platform. The data obtained from the high-throughput sequencing were processed to construct a training data set for the pY class antibody prediction model, referred to as the data set.

[0087] The data obtained from high-throughput sequencing cannot be used immediately for training the antibody-like prediction model. It needs to be processed first and converted into a complete DNA sequence. On this basis, a data set for training the antibody-like prediction model is constructed, including preliminary processing and denoising, as follows:

[0088] 1. Preliminary processing of high-throughput sequencing data;

[0089] (1) Splicing forward and reverse sequencing results on the same sequence: Use flash software to splice two sequencing results on the same sequence that are successfully paired. After completing this operation, a sequence with the same length as the sequencing sequence is obtained;

[0090] (2) Determine the direction of the spliced ​​sequence; the direction from the 5' end to the 3' end of the wild-type sequence is defined as the forward direction, and the reverse complementary sequence is defined as the reverse direction. The direction of the sequence after each round of panning is determined;

[0091] (3) performing a reverse complement operation on the reverse sequence; if the sequence obtained by splicing is a reverse sequence, performing a reverse complement operation on it to adjust it to a forward sequence;

[0092] (4) Quality control: Use the fastq_quality_filter sub-tool in the fastx_toolkit software to perform quality control on the spliced ​​sequences. The command adapted to quality control is: "fastq_quality_filter-q20-p90-iinput.fq-o out.fq-Q 33"; this command means: only retain DNA sequences with a single base quality control value greater than 20 and a base ratio of not less than 90% in a single sequence that meets this quality control value;

[0093] (5) DNA sequence is translated into amino acid sequence; DNA sequence is translated into amino acid sequence according to the triplet codon principle; it should be noted that in order to preserve the sequence information to the greatest extent, the stop codon is also translated into the corresponding "translation operation"; the translation rule of the stop codon is defined as: the codon TAG is translated into q; the codon TAA is translated into x; the codon TGA is translated into w;

[0094] (6) Calculate the number of amino acid sequences; calculate the number of each amino acid sequence.

[0095] After completing the high-throughput sequencing data processing of the SH2 mutant library and each panning round, a DNA sequence with a length of 240 and an amino acid sequence with a length of 80 were obtained. In order to conveniently represent the panning rounds, the original mutant library is represented as Round0, and the panning from the first round to the fourth round is represented as Round1 to Round4. After the SH2 mutant library and the sequencing data of each SH2 panning round were processed, the statistical results of the SH2 mutant sequences obtained are shown in Table 2.

[0096] Table 2 Statistical results of SH2 mutants after sequencing data processing of mutation library and panning library

[0097] Round0 Round 1 Round 2 Round 3 Round4 Number of protein sequence types 823,190 749,341 522,458 131,107 49,211 Protein sequence copy number 1,571,441 1,582,204 1,791,148 4,509,187 3,331,589 Average copy number of protein sequence 1.91 2.11 3.43 34.40 67.70

[0098] The average copy number of the protein sequence in Table 2 is obtained by dividing the copy number of the protein sequence by the number of protein sequence species. The average copy number of the protein can reflect whether an advantageous mutant (a mutant that is more inclined to bind to pY) appears during the bio-panning process. By observing the data in Table 2, it can be seen that as the bio-panning proceeds, the number of protein sequence species gradually decreases, and the copy number of the protein sequence shows an increasing trend. The average copy number remains basically unchanged between Round 0 and Round 2 (Table 2), and gradually increases between Round 2 and Round 4, indicating that the pY-type antibody sequence gradually gains a growth advantage between Round 2 and Round 4.

[0099] 2. Denoising

[0100] The data denoising process is performed according to the following standards:

[0101] (1) Delete the sequence containing the stop codon;

[0102] (2) Delete amino acid sequences whose copy number is less than 3 and not present in all panning rounds.

[0103] According to the above denoising criteria, the data in Round 0 to Round 4 were denoised. After denoising, the statistical results of the SH2 mutant library and the SH2 mutants in each panning round are shown in Table 3.

[0104] Table 3. Data statistics of SH2 mutants after denoising in the mutation library and panning library

[0105] Round0 Round 1 Round 2 Round 3 Round4 Number of protein sequence types 87,334 101,369 97,991 49,436 23,514 Protein sequence copy number 635,797 724,718 1,239,083 4,400,205 3,299,612 Average copy number of protein sequence 7.28 7.15 12.64 89.01 140.33

[0106] Comparing the data in Table 2 and Table 3, we can see that the changes in the average copy number of protein sequences before and after denoising are strongly correlated (Pearson correlation coefficient: 0.99), that is, both show that as the panning rounds proceed, the number of sequence types in the library gradually decreases, and the average copy number of protein sequences gradually increases. The average copy number basically does not change between Round 0 and Round 2, and gradually increases in the rounds after Round 2, reflecting the success of the denoising process.

[0107] The amino acids at the eight mutation sites of the SH2 mutant sequences before and after the two rounds of panning were extracted and sequentially connected to form a peptide segment of eight amino acids in length for the construction of an antibody-like prediction model data set.

[0108] 3. Construction of antibody-like prediction model dataset

[0109] The antibody-like prediction model of the present invention includes a binary classification prediction model and a regression prediction model. The descriptions of these two prediction models and their data sets are as follows:

[0110] Binary classification prediction model and its data set: The binary classification prediction model is used to predict the probability of specific binding between SH2 mutants and pY peptides; its data set is constructed based on the growth amplitude difference data between two rounds.

[0111] Antibody-like regression prediction model: The antibody-like regression prediction model is used to predict the changes in the copy number of SH2 mutants; its data set is constructed based on the growth ratio data between two rounds.

[0112] A higher score in the mutant binary classification prediction model indicates that the mutant has a higher probability of specific binding to pY; a higher score in the mutant regression prediction model indicates that the increase in the mutant copy number between rounds is high; mutants with two high scores are predicted to be pY potential class antibodies with higher affinity.

[0113] In addition, the literature (Li, S., et al. (2021). "Revisiting the phosphotyrosine bindingpocket of Fyn SH2 domain led to the identification of novel SH2superbinders." Protein Science 30(3):558-570) The OD450 signal of the library was measured before and after each round of panning. The change of OD450 signal of SH2 mutant library during bio-panning is shown in Figure 3As shown in the figure, the OD450 signal between Round0 and Round2 basically did not change; the OD450 signal between Round2 and Round4 gradually increased, corresponding to the change in the average copy sequence in Tables 1 and 2. When training the antibody-like prediction model, considering that only data from two rounds are needed to train the model, and because the dominant sequence changes significantly between Round2 and Round4 during the bio-panning process, the sequences in Round2 and Round4 were selected as the training data for the antibody-like prediction model.

[0114] (1) Construction of binary classification prediction model dataset

[0115] The training data set for the binary classification prediction model includes two columns of data: an amino acid sequence of length 8 (sequence column) and a sequence label column (difference column, represented by 0 or 1); 1 in the label column indicates that the mutant sequence of SH2 can bind to the pY modified peptide, and vice versa 0. Follow the steps below to define the labels of the sequences in the training data (i.e. the values ​​in the difference column):

[0116] 1) First, calculate the ratio of the sequence copy number of the SH2 mutant sequence corresponding to the amino acid sequence in Round 4 and Round 2 to the sum of the copy numbers of all mutants in the current round;

[0117] 2) Calculate the difference between the proportion of the SH2 mutant sequence corresponding to the amino acid sequence in Round 4 and the proportion in Round 2, that is, the proportion of the SH2 mutant sequence corresponding to the amino acid sequence in Round 4 minus the proportion of the SH2 mutant sequence in Round 2;

[0118] 3) After calculating the difference between the proportion of the SH2 mutant sequence corresponding to a certain amino acid sequence in Round 4 and the proportion in Round 2, define the label of the sequence according to the following standards:

[0119] If the difference in the proportion of the amino acid sequence in the two rounds of panning is greater than 0, the difference value of the sequence is defined as 1, indicating that the mutant sequence can bind to the pY peptide; otherwise, it is 0, indicating that the mutant sequence cannot specifically bind to the pY peptide.

[0120] The processed sequences were labeled according to the above method. There were 102,746 sequences in total, of which 8460 sequences had a difference value greater than 0, which were positive samples; the remaining 94,286 sequences had a difference value less than or equal to 0, which were negative samples. The positive and negative samples were selected from the maximum and minimum values ​​of the proportion value to the middle, so that the number of positive and negative samples was equal. In order to ensure the balance of the number of positive and negative samples in the data set, all positive samples were taken out and included in the data set of the antibody-like prediction model; 8461 sequences were taken upstream from the sequence with the smallest difference in the negative sample data and included in the data set of the antibody-like prediction model. In this way, the training data set of the binary classification prediction model was obtained.

[0121] In order to view the differences in features between the positive sample data set and the negative sample data set, the present invention draws a Two-Sample-Logo (TSL) diagram of the positive and negative sample data sets (P<0.05, t-test with Bonferroni correction), as shown in Figure 4 As shown. Figure 4 It can be seen that there are significant characteristic differences between positive samples that can specifically bind to the pY peptide and negative samples that cannot bind to the pY peptide, with the maximum difference value being 59.9%. In the positive samples, glutamic acid (E) is significantly enriched at mutation position 1 and mutation position 2; serine is significantly enriched at mutation position 3; valine and alanine are significantly enriched at mutation position 6; lysine is significantly enriched at mutation position 7; and leucine is significantly enriched at mutation position 8. In the negative samples, threonine is significantly enriched at mutation position 4; serine is significantly enriched at mutation position 6; and lysine is significantly enriched at mutation position 8. The obvious difference between positive and negative samples reflects the predictability of the antibody-like.

[0122] In order to better test the predictive ability of the binary classification prediction model of antibodies, 4 / 5 of the binary classification prediction model data set was randomly selected as a cross-validation dataset; the remaining 1 / 5 was used as an independent test set (Table 4).

[0123] Table 4. Statistics of binary classification prediction model dataset

[0124]

[0125] (2) Construction of regression prediction model dataset

[0126] The training data set of the regression prediction model includes two columns of data: an amino acid sequence of length 8 (sequence column) and a label column of the sequence (log(ratio) column); the values ​​in the log(ratio) column are calculated according to the following steps:

[0127] 1) First, calculate the proportion of the sequence copy number of the SH2 mutant sequence corresponding to the amino acid sequence in Round 4 and Round 2 to the total copy number in each round;

[0128] 2) Calculate the ratio of the sequence proportion of the SH2 mutant sequence corresponding to the amino acid sequence in Round4 to the sequence proportion in Round2, and perform a log10 operation; a value greater than 0 means that the sequence proportion increases from Round2 to Round4, and vice versa, it means that the sequence proportion decreases from Round2 to Round4. When calculating the ratio, in order to avoid the situation where the denominator is 0, perform the following processing: when the denominator is 0, assign it to 10 and recalculate. That is, when the proportion value is 0, set the copy number of the sequence with a copy number of 0 to 10 and recalculate the proportion.

[0129] 3) Taking 0 as the boundary, the label values ​​are divided into two categories according to the log10 operation results. The number of sequences in the smaller category is counted as n, and the number of sequences in the larger category is counted as m. The sequences in the smaller category are included in the regression prediction model data set; the same number of sequences are selected from the larger category and also included in the regression prediction model data set; the selection method is: arrange the m sequences according to the size of the label value, divide them into n parts, randomly select a sequence from each part, and finally get n sequences, which are included in the regression prediction model data set.

[0130] After calculation, there are 2253 sequences with log(ratio) greater than 0 in Round2 and Round4, and the remaining 100,493 sequences have log(ratio) less than 0. Based on this, the regression prediction model data set is constructed, as follows:

[0131] ① All sequences with log(ratio)>0 are included in the regression prediction model data set;

[0132] ②Sequences with log(ratio)<0 are arranged in descending order of log(ratio) values ​​and divided into 2253 parts. A sequence is randomly selected from each part, and a total of 2253 data are included in the data set of the regression prediction model; thus, the data set used to build the regression prediction model is obtained.

[0133] The regression model dataset was divided into a training set and an independent test set with a ratio of 4:1. The division method was as follows: the dataset was sorted in descending order according to the value of the log(ratio) column, with 5 sets as one set; one sequence was taken from each set as an independent test set; the rest was the training set (Table 5).

[0134] Table 5. Antibody-like regression prediction model data set statistics

[0135] Training set Independent test set 3606 904

[0136] Example 2 Construction of class 2 antibody prediction model and its performance evaluation

[0137] Deep learning is a type of function in machine learning. The present invention ultimately searches for pY class antibodies that can specifically bind to pY peptide segments with high affinity based on a deep learning algorithm, and predicts pY class antibodies based on the changes in various mutants during the bio-selection process.

[0138] Deep learning is mostly in the form of multi-layer neural networks. Most neural networks are basically composed of the following core principles:

[0139] (1) Linear units and nonlinear units are used alternately, which are often called "layers".

[0140] (2) Use the chain rule (i.e. back propagation) to update the parameters of the network.

[0141] Convolutional Neural Network (CNN) is a commonly used deep learning algorithm. It is a feedforward neural network composed of multiple pooling layers and convolution layers. The basic structure of CNN consists of input layer, convolutional layer, pooling layer, fully connected layer and output layer. The input layer is used to receive training data for model training; the data used for model training is often converted into a digital matrix. The function of the convolution layer is to extract the features present in the input data. The pooling layer can sample the features extracted by the convolution layer, which greatly reduces the redundancy of the data and reduces the problem of overfitting of the model. The function of the fully connected layer is to summarize the information processed by the previous network layers and finally output the results through the output layer. There is usually an activation function in the output layer; the activation function can map the output result to a specified range.

[0142] 1. Construction of antibody-like prediction model

[0143] The antibody-like prediction model of the present invention includes a binary classification prediction model and a regression prediction model, both of which are constructed based on CNN, and the CNN network structures of the two models are completely consistent. Specifically, the present invention uses specific sequence feature encoding combined with CNN network architecture to complete the construction of the prediction model and perform performance testing. The network architecture used in the present invention is shown in FIG. Figure 5 shown.

[0144] The CNN network architecture used is described as follows:

[0145] (1) Input layer: The input layer of the feature vector generated after feature encoding.

[0146] (2) Convolutional layer: Contains four convolutional sublayers and two maximum pooling layers; each convolutional sublayer has 128 convolution kernels of sizes 1, 3, 9, and 10. The two maximum pooling layers are located behind the third and fourth convolutional sublayers, respectively.

[0147] (3) Fully connected layer: This layer consists of four fully connected layers, two maximum pooling layers, and one global average pooling layer. The combination is as follows: the first two fully connected layers are followed by a maximum pooling layer; the third fully connected layer is followed by a global average pooling layer; and the last fully connected layer is at the last position. In order, the number of units in the fully connected layers are 64, 32, 8, and 1, respectively.

[0148] (4) Output layer: This layer contains one neuron that outputs a decimal. It is activated by the sigmoid activation function and outputs a value between 0 and 1.

[0149] The present invention uses One-Hot encoding to convert the amino acid sequence in the data set into a digital matrix that can be recognized by the CNN model. In One-Hot encoding, 20 amino acids and are encoded as 20-dimensional binary vectors. In the vector corresponding to the amino acid, the element related to the amino acid is marked as 1, and the other elements are marked as 0. For example, "A" is represented by "100000000000000000000", and "C" is represented by "01000000000000000000".

[0150] 2. Prediction model performance evaluation

[0151] (1) Performance evaluation indicators of binary classification prediction models

[0152] In the binary classification task, the sequence can be divided according to the combination of the true category and the predicted category, namely: True Positive (TP), False Positive (FP), True Negative (TN), False Negative (FN). Based on this, the performance evaluation index values ​​such as sensitivity (SN), specificity (SP), overall accuracy (ACC) and Matthew correlation coefficient (MCC) can be further calculated. The present invention also calculates the area under the ROC curve value (Area Under Curve, AUC) to evaluate the overall performance of the binary classification prediction model.

[0153] (2) Performance evaluation indicators of regression prediction models

[0154] The commonly used evaluation indicators of regression models are: Mean Absolute Error (MAE), Mean Square Error (MSE), and goodness of fit (R-Squared; R2).

[0155] The present invention uses a convolutional neural network as a prediction model algorithm and One-Hot as an encoding method for input data to construct a binary classification prediction model. The role of the binary classification prediction model is to predict whether a given sequence can specifically bind to pY. The prediction model is trained using a cross-validation set with a ten-fold cross-validation training method, and then the performance of the trained prediction model is independently tested using an independent data set. The performance statistics of the prediction model on the ten-fold cross-validation and independent test are shown in Table 6 below.

[0156] Table 6. Performance statistics of binary classification prediction model

[0157] SN SP ACC MCC AUC Ten-fold cross validation 0.899±0.015 0.957±0.0127 0.929±0.007 0.859±0.014 0.966±0.005 Independent testing 0.891±0.007 0.963±0.004 0.927±0.002 0.857±0.004 0.968±0.002

[0158] A regression model was constructed using a network architecture consistent with the binary classification model. The regression model was trained using the training set, and then the trained regression model was independently tested using an independent dataset. The results are shown in Table 7.

[0159] Table 7. Regression model performance statistics

[0160] <![CDATA[R 2 ]]> MAE MSE 0.7704 0.1911 0.0786

[0161] In order to further determine the accuracy of the regression model prediction after training, the present invention calculates the correlation between the predicted score (Pre_log(ratio)) of the independent test set and the true value (Log(ratio)). The scatter plot of the relationship between the predicted score of the independent test set and the true value is shown in Figure 6 .Depend on Figure 6 It can be seen that the model's fitting ability (R 2 ) is 0.7704, and the overall fitting ability is excellent.

[0162] Example 3 A method for predicting antibody-like products that bind to biomacromolecule modifications based on protein domains

[0163] After the antibody-like prediction model is constructed and trained, the antibody-like prediction model can be used to explore sequence features and predict antibody-like modified with biomacromolecules. After the prediction is completed, the corresponding antibody-like sequence is obtained according to the sequence features, and the predicted antibody-like sequence is evaluated using protein enzyme-linked immunosorbent assay (Protein ELISA) and affinity determination experiments.

[0164] The process of exploring sequence features and predicting antibody classes based on sequence features is as follows:

[0165] In the process of bio-panning the SH2 mutant library using pY peptides, the proportion of mutants that are more inclined to specifically bind to pY in each round of panning tends to show a gradually increasing trend. In order to find the sequence characteristics of pY-type antibodies, we first find out the amino acids that have an enrichment trend at each mutation position during the panning process; these enriched amino acids are arranged and combined according to the order of mutation positions to generate an amino acid sequence to be predicted with a length of 8.

[0166] Among them, the method for determining the enriched amino acids is as follows:

[0167] ① Use the binary classification model score and regression model to score all the denoised sequencing sequences;

[0168] ②Select sequences with scores greater than the threshold; threshold: binary classification prediction model score greater than 0.5; regression prediction model score greater than the score of Fyn SH2 wild type;

[0169] ③ Count the proportion of amino acids at each mutation position;

[0170] ④ The amino acids with a proportion greater than 0.05 at each position (High frequency amino acid) and the enriched amino acids (TSL enriched amino acid) in the TSL map are defined as enriched amino acids.

[0171] Generation of sequences to be predicted: After determining the enriched amino acids at each mutation position, the sequences to be predicted are arranged and combined in order of the mutation positions to generate sequences to be predicted. The specific rules are as follows: one of the 8 positions is allowed to be a common amino acid (20 species), and the remaining 7 positions must be enriched amino acids. Based on the above method, the present invention generated 5,664,001 sequences. It should be noted that the antibody-like model scores are higher than the Fyn SH2 wild type sequences, so the generated sequences to be predicted are scored again using the trained antibody-like prediction model. It can be predicted that the scores of 7466 sequences are higher than the wild type score (wild type binary classification model score: 0.0012; regression model score: 0.2787), of which there are 4971 sequences to be predicted that do not appear in the sequencing data.

[0172] In order to select high-affinity pY class antibodies (the lower the scores of the two models, the lower the affinity), the above 7466 sequences were filtered. The filtering criteria were: the prediction score of the binary classification model > 0.9 and the prediction score of the regression model > 1. 293 sequences were screened out and called candidate class antibodies. It is worth noting that the superbinders reported in the literature (Kaneko, T., et al. (2012) "Superbinder SH2 domains act as antagonists of cell signaling." Science signaling 5(243):ra68.; Li, S., et al. (2021). "Revisiting the phosphotyrosinebinding pocket of Fyn SH2 domain led to the identification of novel SH2superbinders." Protein Science 30(3):558-570.), such as SH2 superbinder, V3, V10, V13 and V24, which are all present in the potential pY class antibodies. In addition, 108 candidate class antibodies do not exist in the sequencing data.

[0173] Since there are still a large number of candidate antibodies, we further conducted sequence logo conservation analysis on the candidate antibodies. The results are as follows: Figure 7 As shown. Figure 7 It can be seen that the P4 position is the least conservative, while the other positions are relatively conservative. Since the preference of the amino acid at the P4 position cannot be determined, the 293 candidate antibodies were split according to the amino acid at the P4 position to obtain 9 groups. Sequence Logo analysis was performed again for each group of candidate antibodies. The results are as follows Figure 8 As shown. Figure 8 It can be seen that the amino acids at each position of each group of sequences are relatively conservative, and it is believed that the sequences with conservative amino acids at each position are more likely to be antibody-like. The definition of conservative amino acids is as follows: the one or two amino acids with the highest values ​​at the mutation position. The most conservative amino acids at each mutation position in each group are determined, and the most conservative amino acids are combined in sequence to form a sequence that is the predicted antibody-like.

[0174] Based on the above method, the present invention lists the sequence consisting of 18 conservative amino acids, and randomly selects 4 as comparison from 293 candidate class antibodies at the same time, obtains corresponding class antibody sequence and carries out ELISA and affinity determination experiment according to sequence characteristics. The sequence consisting of 18 conservative amino acids and affinity determination results thereof are summarized as shown in Table 8, and these sequences are finally predicted to be SH2 class antibodies, named M_4* (wherein 4 is the 4th mutation position, * is the amino acid at this position), and 4 SH2 class antibodies (Trim, V13, V50 and V3) reported are included in the predicted SH2 class antibodies. As a comparison, 4 randomly selected from 293 candidate class antibodies are named C_4*, which are called contrast mutants, and their sequence and affinity determination results are as shown in Table 9.

[0175] Table 8. SH2 predicted antibody sequence list

[0176]

[0177] 1 Naming principle: M represents mutant, 4 represents the 4th mutation position, the letter after 4 represents the amino acid at the 4th position, and the number after that represents the ranking number in the group. For example, M_4T1 is the first member of the mutant group with T at the 4th position.

[0178] 2 The KD column shows the KD value between the antibody-like sequence and the peptide pYEEI.

[0179] The "Not Tested (low ELISA value)" in the table means that the value of the sequence in the ELISA test is low, and no further affinity test is performed on it. The reason for not testing is that the affinity of antibody-like sequences with the same sequence characteristics is comparable, and there is no need to test all predicted antibody-like sequences with the characteristics, which is meaningless, so a few sequences were not tested.

[0180] Table 9. Randomly selected antibody-like sequences

[0181]

[0182] 1 Naming principles: C represents the mutant used for comparison, 4 represents the 4th position, the letter after 4 represents the amino acid at the 4th position, and the number after that represents the serial number.

[0183] The present invention conducted ELISA experiments on the predicted class antibodies (shown in Table 8) and the comparative class antibodies (shown in Table 9) using the phosphorylated peptide pYEEI, respectively, with the wild type (WT) and mutant Trim as controls. The ELISA experimental results of the predicted class antibodies and the comparative class antibodies and the phosphorylated peptide pYEEI are shown in Figure 1. Fig. 9 and Fig.10 As shown. For mutants with ELISA signals higher than WT, Biacore was further used to determine the affinity of the predicted antibody-like peptides to pYEEI and GGpYGG. The affinity determination results of the predicted antibody-like peptides to pYEEI and GGpYGG are shown in Figure 2. Fig.11 and Fig.12 As shown in Table 8, Table 9, Fig.11 and Fig.12 The results show that the affinity of the predicted antibody-like is much higher than that of the mutant used for comparison, and is close to the affinity of the reported antibody-like. In addition, the affinity of the antibody-like used for comparison is mostly higher than that of WT. The above results show that the algorithm for predicting antibody-like based on artificial intelligence is effective and reliable, and the method for predicting antibody-like combined with biomacromolecule modification described in the present invention can obtain high-affinity antibody-like.

[0184] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for predicting antibody-like binding to biomacromolecule modification based on protein domains, characterized in that: The following steps are involved: S1. Based on the protein domain, analyze the amino acid sites that may bind to biomacromolecule modification and mutate them, obtain the mutant library and then perform panning; S2. Perform high-throughput sequencing on the mutant library and the library obtained by panning, process the sequencing data, obtain mutant sequences and construct an antibody-like prediction model data set; the antibody-like prediction model data set includes an antibody-like binary classification model and a regression model data set; S3. constructing an antibody-like binary classification prediction model and a regression prediction model based on a convolutional neural network, and using the constructed binary classification model and regression model data sets to train the constructed binary classification prediction model and regression prediction model respectively; S4. Use the trained binary classification prediction model and regression prediction model to score the mutant sequences obtained in step S2, retain the mutant sequences with higher scores than the wild type, count the enriched amino acids at each mutation position of the mutant sequence, arrange and combine the amino acids according to the order of the mutation positions, and generate a sequence to be predicted; S5. Calculate the conservation of the amino acid at each mutation position in the sequence to be predicted, and group the sequences to be predicted according to the amino acid at the least conserved mutation position; recalculate the conservation of the amino acid at each mutation position of the sequence to be predicted in each group, determine the most conserved amino acid at each position in each group, combine the most conserved amino acids in sequence, and the formed sequence is the predicted antibody-like protein.

2. The method according to claim 1, characterized in that: The processing of the sequencing data in step S2 includes preliminary processing and denoising processing; wherein the preliminary processing is: splicing the DNA sequence obtained by high-throughput sequencing data conversion and performing quality control, setting the quality control value of a single base to be not less than 20, retaining DNA sequences in which the base ratio that meets this quality control value is not less than 90%, and translating the retained DNA sequences into amino acid sequences; the denoising processing is: deleting the amino acid sequence with a copy number less than 3 and not present in all panning rounds; after processing, the mutant sequence is obtained.

3. The method according to claim 1, characterized in that: The binary classification prediction model data set in step S2 includes two columns of data: a data column and a label column; the data column is an amino acid sequence composed of amino acids at the mutation positions of the mutant sequence, and the value of the label column is defined as 0 or 1; the definition method of the label value of the amino acid sequence is: (1) Calculate the proportion of the copy number of the amino acid sequence in the total copy number of the sequence in each panning round; (2) Calculate the difference in the proportion of the amino acid sequence in the two rounds of data before and after the panning; define the amino acid sequence as follows based on the difference in proportion: If the proportion of the amino acid sequence after panning is greater than the proportion before panning, the label value of the amino acid sequence is defined as 1, which is a positive sample; otherwise, it is 0, which is a negative sample. When the number of positive and negative samples is unequal, select from the maximum and minimum values ​​of the proportion value to the middle to make the number of positive and negative samples equal. The regression model data set includes two columns of data: a data column and a label column; the data column is an amino acid sequence composed of amino acids at the mutation positions of the mutant sequence in sequence; the value of the label column is defined as follows: (1) Calculate the proportion of the amino acid sequence copy number in the total copy number of each round of sequences; (2) Calculate the ratio of the proportion after panning to the proportion before panning, and perform a log10 operation on the ratio; when the proportion is 0, set the copy number of the sequence with a copy number of 0 to 10, and recalculate the proportion; (3) Taking 0 as the boundary, the label values ​​are divided into two categories according to the log10 operation results. The number of sequences in the smaller category is counted as n, and the number of sequences in the larger category is counted as m. The sequences in the smaller category are included in the regression prediction model data set; the same number of sequences are selected from the larger category and also included in the regression prediction model data set; the selection method is: arrange the m sequences according to the size of the label value, divide them into n parts, randomly select a sequence from each part, and finally obtain n sequences, which are included in the regression prediction model data set.

4. The method according to claim 1, characterized in that: The method for constructing the binary classification prediction model and the regression prediction model in step S3 is as follows: (1) Determine feature encoding: input the feature vector generated after feature encoding into the input layer, use the algorithm to test the commonly used protein encoding methods, and evaluate the performance of different feature encodings by cross-validation; select the encoding with the best performance as the feature encoding of the data set; the commonly used protein encoding methods tested by the algorithm include One-Hot encoding; (2) constructing input layer, convolution layer, fully connected layer and output layer models: the difference between the input layer, convolution layer, fully connected layer and output layer models is only the number of convolution layers, and the number of convolution layers of the model is 2 to 6; The convolution kernel size of the first convolution layer is fixed to 1, the convolution kernel size of the remaining convolution layers is fixed to 3, the number of convolution kernels of all convolution layers is fixed to 64, and a dropout layer is added after all convolution layers, and the parameter is fixed to 0.

5. The number of fully connected sublayers in the fully connected layer is fixed to 3, the number of units in the last fully connected sublayer is set to 1, the number of units in the remaining fully connected sublayers is fixed to 32 and a dropout layer is added at the end with a fixed parameter of 0.5, and there is a global average pooling layer before the last fully connected sublayer; The output layer includes one neuron and uses sigmoid as the activation function; the patience of the early stopping strategy is fixed to 100, the batch_size value is fixed to 256, and the epoch parameter is fixed to 1000.

5. The method according to claim 1, characterized in that: In step S4, the amino acids with a proportion greater than 0.05 at each position and the enriched amino acids in the drawn Two-Sample-Logo map are defined as enriched amino acids.

6. A system for predicting antibody-like binding to biomacromolecule modification based on protein domains, characterized in that: It includes a high-throughput sequencing data processing module, an antibody-like prediction model evaluation module, and a prediction sequence output module; The high-throughput sequencing data processing module is used to process high-throughput sequencing data and construct an antibody-like binary classification model and a regression model data set, and submit the data set to the antibody-like prediction model evaluation module; The antibody-like prediction model evaluation module includes a binary classification prediction model and a regression prediction model of the antibody-like, firstly trains the binary classification prediction model and the regression prediction model using the input binary classification model and regression model data sets, then inputs the sequence to be evaluated into the binary classification prediction model and the regression prediction model, and outputs the binary classification prediction model score and the regression prediction model score; The predicted sequence output module screens the sequences according to the input binary prediction model scores and the regression prediction model scores, retains the sequences with scores higher than the wild type, counts the enriched amino acids at each mutation position of the sequence, arranges and combines the amino acids according to the order of the mutation positions, generates the sequence to be predicted and calculates the conservation of the amino acids at each mutation position in the sequence to be predicted, groups the sequences to be predicted according to the amino acids at the least conserved mutation position; recalculates the conservation of the amino acids at each mutation position of the sequence to be predicted in each group, determines the most conserved amino acid at each position in each group, combines the most conserved amino acids in sequence, forms and outputs the predicted antibody-like sequence.

7. An antibody-like substance that binds to a peptide containing tyrosine phosphorylation modification, characterized in that: The antibody-like protein is predicted by the method of claim 1; the amino acid sequence of the antibody-like protein is shown in any one of SEQ ID NOs. 2 to 11.

8. A DNA sequence encoding the antibody of claim 7.

9. A vector containing the DNA sequence of claim 8.

10. Use of the antibody of claim 7 in the preparation of a product for detecting peptides containing tyrosine phosphorylation modifications.