Method and system for improving antibody affinity based on molecular dynamics and machine learning

Abstract

The present invention discloses a method and system for improving antibody affinity based on molecular dynamics and machine learning, and belongs to the field of computer software technology. The method of the present invention uses molecular dynamics to accurately locate the binding site of the antibody, thereby screening out the best amino acid mutation site, and constructing a variant library of amino acid saturation mutation at the best amino acid mutation site, and calculating the change (value) of the binding free energy of the mutant complex as the evaluation basis; using the amino acid descriptor as a feature and the value as a label, training the machine learning model, and screening out the model with the best prediction performance; the model is used to predict and screen out mutants with lower values ​​(i.e., high affinity potential). The high-affinity 1-7 single-domain antibody obtained by the method of the present invention has a significantly improved affinity of 2.98-5.04 times, which effectively verifies the practicality and efficiency of the present invention and opens up a new path for the directed evolution of single-domain antibodies.

Description

Technical Field

[0001] The present invention belongs to the field of computer software technology, and specifically relates to a method and system for improving antibody affinity based on molecular dynamics and machine learning. Background Art

[0002] In the field of biomedicine, especially in antibody design and optimization, with the deepening of genomics and proteomics research, scientists have a deeper understanding of the structure and function of natural proteins. This accumulated information provides a rich resource for discovering proteins with specific functions. However, although natural proteins have a variety of biological activities, it is often difficult to meet all application requirements, especially in the development of antibody drugs, which requires extremely high binding affinity between antibodies and target antigens. Traditional methods for improving antibody affinity, such as ribosome display, yeast two-hybrid, phage display antibody library, etc., although to a certain extent, have achieved the directed evolution of antibodies, there are many limitations. First, these technologies are difficult to construct a mutation library that comprehensively covers all possible mutation sites and amino acid substitutions, resulting in the omission of potential optimization schemes. Secondly, the construction of antibody libraries and their screening process are not only time-consuming and labor-intensive, but also limited by the difficulty of target protein expression and the stability of the in vitro screening environment. Furthermore, as the size of the library increases, although the possibility of including the best mutants increases, the existing screening technology is difficult to process such a large library within a reasonable time, thereby limiting the efficiency of functional improvement. Therefore, it is particularly important to explore more efficient and accurate antibody affinity optimization methods.

[0003] Compared with the traditional method of improving antibody affinity, the method of obtaining high-affinity antibodies by computer-aided design can save experimental time; single-point and combined virtual mutations can be performed on all sites of the antibody binding domain, but this method requires high computing resources, and the calculation accuracy and efficiency need to be improved. In recent years, the rapid development of molecular dynamics simulation and machine learning technology has brought new hope to this field. Molecular dynamics simulation can accurately simulate the interaction between protein and ligand, locate key binding sites, and provide a scientific basis for mutation design. At the same time, molecular dynamics simulation can explore the flexibility of proteins to a greater extent and deal with entropy components more clearly, so as to estimate binding affinity with higher accuracy. Compared with traditional experimental methods, molecular dynamics simulation can predict possible results before the experiment begins, making the experiment more directional, which helps researchers identify potential optimization points in advance and reduce the blindness of the experiment. Therefore, molecular dynamics simulation can evaluate multiple possible mutation schemes in a short time, greatly improving the efficiency of antibody affinity optimization. However, existing molecular dynamics simulations involve the calculation of a large number of atomic and molecular interactions, which requires extremely high computing resources and is very time-consuming to obtain accurate binding free energy estimates. Moreover, when our mutation library is very large, the estimation of the binding free energy of mutants by molecular dynamics simulation requires a lot of manpower, material resources, financial resources and time. Machine learning technology can not only efficiently process large-scale data sets and extract features that affect antibody affinity, providing strong data support for antibody affinity optimization, but also learn the relationship between protein sequence, structure and function through big data analysis, predict the effect of mutations on affinity, and thus guide the design of mutation libraries. By constructing a predictive model, machine learning can evaluate a large number of antibody sequences in a short time and quickly screen candidate antibodies with high affinity. However, the performance of machine learning models is highly dependent on the quality and quantity of training data. In antibody affinity prediction, high-quality experimental data is often difficult to obtain and there is a data imbalance problem. Molecular dynamics simulations can show in detail the interaction process between mutant antibodies and their antigens, including changes in binding free energy, which is crucial for understanding how mutations affect the binding affinity of antibodies and antigens. Therefore, it is very promising to combine molecular dynamics with machine learning techniques to improve the binding affinity of antibodies.

[0004] Improving antibody affinity involves knowledge from multiple fields such as biology, chemistry, and computer science. It is a challenge to effectively integrate knowledge from these fields into machine learning models. It is necessary and meaningful to explore how to determine mutation sites, build variant libraries, and quickly and accurately screen out high-affinity antibodies from a large variant library. Summary of the invention

[0005] In view of the fact that it is very difficult to obtain ideal high-affinity antibodies, and that they have low efficiency, long time, and the required mutation library is too large, the present invention aims to provide a method and system for improving antibody affinity based on molecular dynamics and machine learning.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention provides a method for improving antibody affinity based on molecular dynamics and machine learning, comprising:

[0008] S1, obtain the amino acid sequence of the antibody and preprocess to obtain the best model for antigen-antibody docking;

[0009] S2, based on the best model of antigen-antibody docking, the simulation tool is used to analyze and combine the molecular dynamics calculation method to obtain the best amino acid mutation site of the antibody;

[0010] S3, randomly inserting all possible amino acids at the optimal amino acid mutation site of the antibody to generate a variant library;

[0011] S4, extracts the feature vectors of mutants in the variant library as input, uses the change in binding free energy as a label, and inputs it into a pre-trained machine learning model to predict the change in binding free energy of mutants in the variant library and screen potential high-affinity candidate antibodies.

[0012] The S1 includes:

[0013] S11, sequencing the antibody by protein sequencing technology to obtain amino acid sequence information of the antibody, wherein the sequence information is composed of sequence information of a plurality of amino acid types connected end to end;

[0014] S12, inputting the amino acid sequence of the antibody obtained in S11 and the amino acid sequence of the known antigen into the AlphaFold2 online server to predict the full-length structure of the antibody and the antigen, selecting the structure with the highest ranking based on the score as the optimal structure, and determining the position and sequence of the CDRs region of the antibody through the IgBlast tool platform, wherein the optimal structure information is composed of the three-dimensional coordinates of the atoms of several amino acids connected head to tail;

[0015] S13, the optimal structures of the antibody and antigen predicted in S12 are input into the ClusPro 2.0 online server for antigen-antibody molecular docking, and the docking model with the highest ranking is selected as the optimal model for antigen-antibody docking.

[0016] In S2, the optimal amino acid mutation site of the antibody obtained by combining the molecular dynamics calculation method includes:

[0017] S21, based on the best model of antigen-antibody docking, uses simulation tools to perform structural analysis of antigen-antibody docking, identifies potential mutation sites in antibodies, performs single-point saturation mutagenesis on these mutation sites, and obtains antigen-antibody complexes after single-point mutation;

[0018] S22, calculate the change in binding free energy of the antigen-antibody complex after single-point mutation, and select single-point mutants with a change in binding free energy less than -1 kcal / mol;

[0019] S23, combining the single point mutants to generate double mutants, calculating the change in binding free energy of the antigen-antibody complex after the double mutation, and determining the double point mutant with the lowest change in binding free energy as the best double point mutant;

[0020] S24, combining the best double-point mutant obtained in S23 with the single-point mutant selected in S22 to generate a triple-point mutant, calculating the change in binding free energy of the antigen-antibody complex after the triple-point mutation, and determining the triple-point mutant with the lowest change in binding free energy as the best triple-point mutant. The amino acid mutation site of the triple-point mutation is the best amino acid mutation site for the antibody.

[0021] Furthermore, the simulation tool is any one or a combination of InterProSurf, PyMol or Rosetta protein simulation tools; the change in binding free energy is calculated using the following formula:

[0022] (1)

[0023] in: represents the gas constant; represents absolute temperature; represents the dissociation constant of the complex; represents the binding free energy;

[0024] (2)

[0025] in, represents the binding free energy of the antigen-antibody complex after the antibody is mutated; represents the binding free energy of the original antigen-antibody complex; It represents the change in binding free energy. A positive value indicates that the binding free energy of the complex after mutation increases and the affinity decreases; a negative value indicates that the binding free energy of the complex after mutation decreases and the affinity increases.

[0026] In S4, the pre-trained machine learning model includes:

[0027] S41, select 200 mutants from the variant library as the first round library, calculate the change in binding free energy of the mutants, define the feature vector of each amino acid by using the amino acid descriptor, define the feature vector of the mutant by connecting the feature vector of the amino acid at each mutation site, use the feature vector of the mutant in the first round library as input, combine the free energy change as a label to train the machine learning model, perform regression fitting on the feature vector of the mutant and the label binding free energy change, and obtain the first round optimal machine learning model according to the model evaluation index;

[0028] S42, using the first round of optimal machine learning model to predict the binding free energy changes of mutants in the variant library whose binding free energy changes were not calculated, arranging the predicted binding free energy changes of the mutants from low to high, selecting the top 50 mutants, calculating the binding free energy changes of the mutants, adding the binding free energy changes of these 50 mutants and their corresponding feature vectors to the training data, retraining the machine learning model with the updated training data, performing regression fitting, and obtaining the second round of optimal machine learning model based on the model evaluation indicators.

[0029] S43, using the second round of optimal machine learning model to predict the binding free energy changes of the remaining mutants in the variant library whose binding free energy changes have not been calculated, the binding free energy changes of all mutants are arranged from low to high, and the top five mutants are selected as potential high-affinity candidate antibodies.

[0030] Furthermore, the amino acid descriptor defines a feature vector of each amino acid, and the amino acid descriptor is the amino acid descriptor with the best screening effect in a comparative experiment.

[0031] Furthermore, the machine learning model adopts a support vector machine, a Gaussian process or a random forest combined with a particle swarm optimization model.

[0032] Furthermore, the machine learning model is a particle swarm optimization combined with a support vector machine model.

[0033] Furthermore, the model evaluation index includes the absolute coefficient , mean absolute error and RMS error ;

[0034] The absolute coefficient Calculated by formula (3):

[0035] (3)

[0036] The mean absolute error Calculated by formula (4):

[0037] (4)

[0038] The root mean square error Calculated by formula (5):

[0039] (5)

[0040] in: Represents the model's prediction data; Represents real data; is the real data mean; is the index; n is the number of samples.

[0041] Furthermore, the method also includes the steps of cloning the screened potential high-affinity candidate antibodies, constructing plasmid transfection for protein expression, verifying the affinity of the expressed protein by affinity determination experiments, and screening antibodies with high affinity.

[0042] The present invention provides a system for improving antibody affinity based on molecular dynamics and machine learning, comprising:

[0043] Data preprocessing and model building module, used to obtain the amino acid sequence of the antibody and preprocess to obtain the best model for antigen-antibody docking;

[0044] The mutation site identification and variant library generation module uses simulation tools and molecular dynamics calculation methods to identify the best amino acid mutation sites of antibodies and generate variant libraries based on the optimal model of antigen-antibody docking;

[0045] The machine learning model training and prediction module is used to predict the change in binding free energy of mutants in the variant library and screen potential high-affinity candidate antibodies.

[0046] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method for improving antibody affinity based on molecular dynamics and machine learning are implemented.

[0047] The present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for improving antibody affinity based on molecular dynamics and machine learning are implemented.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The method for improving antibody affinity based on molecular dynamics and machine learning provided by the present invention combines molecular dynamics calculation method with machine learning technology to construct a new antibody design optimization framework, which overcomes the time-consuming and labor-intensive problem of predicting mutant binding free energy by molecular dynamics calculation method and the dependence of machine learning technology on high-quality data sets, and achieves the purpose of quickly and accurately obtaining high-affinity antibodies. Predicting potential optimized mutation sites through molecular dynamics simulation, and combining machine learning technology to learn the relationship between mutation characteristics and binding free energy, so as to accurately construct a small library rich in ideal target protein variants, overcoming the problem that the mutation library in traditional methods is too large and lacks pertinence; using machine learning models to quickly screen mutants, directly screen out sequences with improved characteristics, greatly shorten the screening process and improve the screening efficiency. This efficient screening strategy makes it possible to obtain high-affinity antibodies in a short time.

[0050] Furthermore, by gradually increasing the complexity of mutations (from single to double to triple), we systematically explored the combined effects of multiple mutation sites in antibodies. value, and filter out the values ​​that can significantly reduce The combination of mutations with the highest value (i.e., improving binding affinity) provides a powerful guide for subsequent antibody optimization experiments, combining the accuracy of molecular dynamics with the efficiency of computational chemistry; by iteratively training and optimizing the machine learning model, the prediction of antibody mutants is gradually improved. The prediction accuracy of the value can be improved, thereby more effectively screening mutants with optimization potential.

[0051] The system provided by the present invention adopts a method based on molecular dynamics and machine learning technology to improve antibody affinity, and has significant technical effects such as high efficiency, accuracy, flexibility and innovation. By combining molecular dynamics and machine learning technology, a full-chain antibody optimization process from theoretical prediction to experimental verification is realized. This fusion not only improves the efficiency and accuracy of antibody design, but also brings new ideas and methods to the field of antibody engineering; through automated data collection, preprocessing, model construction, mutation site prediction and variant library generation, as well as intelligent machine learning prediction and experimental verification processes, the automation and intelligence level of antibody design has been greatly improved; these technical effects have jointly promoted the research progress in the field of antibody engineering and provided strong technical support for the research and development of antibody drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of the method for improving antibody affinity based on molecular dynamics and machine learning of the present invention;

[0053] Figure 2Schematic diagram of the three-dimensional structure and docking complex of the c-MET antigen extramembrane binding domain and target antibody in the embodiment of the present invention;

[0054] Figure 3 For the present invention, 200 mutants were selected as the first round library and Rosetta Flex Calculate the change in their binding free energy value;

[0055] Figure 4 The experimental results of modeling using different machine learning models of AA2+PSSM amino acid descriptors in the embodiments of the present invention;

[0056] Figure 5 These are the results of indirect ELISA affinity determination of the 1-7-HWK, 1-7-YWK, 1-7-HWY, 1-7-SWK and 1-7-SWQ antibody mutants in the examples of the present invention. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0058] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0059] Example 1

[0060] like Figure 1 As shown, a method for improving the affinity of an antibody to a target antigen (c-MET) based on molecular dynamics and machine learning is provided, including:

[0061] S1, obtain the amino acid sequence of the antibody and preprocess to obtain the best model for antigen-antibody docking;

[0062] S2, based on the best model of antigen-antibody docking, the simulation tool is used to analyze and combine the molecular dynamics calculation method to obtain the best amino acid mutation site of the antibody;

[0063] S3, randomly inserting all possible amino acids at the optimal amino acid mutation site of the antibody to generate a variant library;

[0064] S4, extracting feature vectors of mutants in the variant library as input, using the change in binding free energy as a label, inputting into a pre-trained machine learning model, predicting the change in binding free energy of mutants in the variant library, and screening potential high-affinity candidate antibodies;

[0065] The S1 includes:

[0066] S11, sequencing the antibody by protein sequencing technology to obtain amino acid sequence information of the antibody, wherein the sequence information is composed of sequence information of a plurality of amino acid types connected end to end;

[0067] In this embodiment, S11 is specifically implemented by obtaining the amino acid sequence information of the target antibody through mass spectrometry protein sequencing technology based on mass-to-charge ratio: VESGGGMVQPGGSLRLSCAASGFTFSNYDMSWVRQAPGKGPEWVSSINSYTYGDGTYYADSVKGRFTISRDNAKNTLYLQMNSLKPEDTALYYCTRGPRATMASDYWGQGTQVTVSS (as shown in SEQ ID NO.1 in the ST.26 standard sequence nucleotide or amino acid sequence table).

[0068] S12, using the AlphaFold 2 online server to predict the full-length structure of the antibody and antigen, selecting the highest-ranked structure based on the score as the best structural information, and determining the CDRs region of the antibody through the IgBlast tool platform, wherein the best structural information is composed of the three-dimensional coordinates of the atoms of several amino acids connected head to tail;

[0069] S13, antigen-antibody molecular docking was performed using the ClusPro 2.0 online server, and the docking model with the highest ranking was selected as the optimal model for antigen-antibody docking. For details, see Appendix Figure 2 shown.

[0070] By the attached Figure 2 The data shows that Figure 2In the figure, (A) is the three-dimensional structure of the c-MET antigen extracellular binding domain, (B) is the three-dimensional structure of the target antibody, (C) is the antigen-antibody docking complex, and (D) is the local three-dimensional structure of the key amino acid sites of the antigen-antibody docking interface. The three-dimensional structure of the c-MET antigen extracellular binding domain shows the overall morphology and key features of the antigen; the three-dimensional structure of the target antibody shows the overall morphology and key structural features of the antibody, including the CDRs region. By comparing the structures of the antigen and antibody, we can further understand the interaction mechanism between them; the antigen-antibody docking complex shows the overall morphology of the antigen and antibody after docking, as well as the interaction interface between them, and we can intuitively see how the antigen and antibody are combined together; the local three-dimensional structure of the key amino acid sites of the antigen-antibody docking interface further shows the detailed information of the antigen-antibody docking interface; through this schematic diagram, we can further understand the conformational changes and stability of the antibody after binding to the antigen. It can be seen that this information is of great significance for understanding the affinity of the antibody and optimizing the antibody design, and verifies the accuracy and reliability of the prediction and docking results in the S1 step.

[0071] The S2 includes:

[0072] S21, using simulation tools to comprehensively analyze the structure of antigen-antibody docking, identify potential mutation sites in antibodies, perform single-point saturation mutagenesis on these mutation sites, and obtain antigen-antibody complexes after single-point mutation;

[0073] In this embodiment, S21 is specifically implemented by using InterProSurf, PyMol and Rosetta protein simulation tools for comprehensive analysis, and finally, 12 consensus mutation sites screened by all three methods are determined: T24, S26, N27, Y28, D29, Y50, T51, Y52, T101, M102, A103, D105. Single-point saturation mutagenesis is performed on these mutation sites to obtain the corresponding single-point mutant antigen-antibody complexes.

[0074] S22, using FoldX and Rosetta to calculate the antigen-antibody complex after single-point mutation Value, select There were 16 single-point mutants with values ​​lower than -1 kcal / mol, and these 16 single-point mutants involved 4 amino acid mutation sites: Thr24, Asn27, Asp29, and Thr51;

[0075] S23, the single-point mutants of the four amino acid residues involved in S22 were combined in pairs to generate double-point mutants, and Rosetta Flex Calculation of the antigen-antibody complex after double point mutation Value, select The value was the lowest, thus D29F / T51W was determined to be the best double-point mutant, and the amino acid mutation sites involved in the double-point mutation were obtained as D29 and T51;

[0076] S24, amino acid residues D29 and T51 involved in the double point mutation obtained in S23 The double-point mutants with values ​​lower than -1 kcal / mol were different from those at T24, S26, N27, Y28, Y50, Y52, T101, M102, A103, and D105. Single point mutants with values ​​below -1 kcal / mol were combined to generate triple point mutants using Rosetta Flex Calculation of the antigen-antibody complex after three point mutations Value, select The value was the lowest, thus determining D29H&T51W&A103R as the best triple-point mutant, thereby determining the three best amino acid mutation sites: D29, T51, and A103.

[0077] The S3 includes:

[0078] A programming language (Python) was used to randomly insert all possible amino acids at the three selected amino acid mutation sites (D29, T51, A103) in the amino acid sequence of the target antibody to generate a variant library.

[0079] The S4 includes:

[0080] S41, select 200 mutants required for the machine learning model from the variant library as the first round library, using Rosetta Flex Computational mutants The amino acid descriptor is used to define the feature vector of each amino acid. The feature vectors of the amino acids at each mutation site are connected to define the feature vector x of the mutant. The feature vectors of the mutants are characterized by various amino acid descriptors shown in Table 1. Benchmark experiments are performed to select the best amino acid descriptor AA2+PSSM. The feature vectors and labels of the mutants in the first round of the library are used. Value training optimization test particle swarm optimization (PSO) machine learning model, feature vectors and labels of mutants The values ​​are regressed and fitted according to the absolute coefficient of the model evaluation index ( ), mean absolute error ( ) and the root mean square error ( ) Select the best particle swarm optimization (PSO) machine learning model to improve the machine learning model to predict mutants The accuracy of the value is used to obtain the optimal machine learning model in the first round.

[0081] (3) (4)

[0082] (5)

[0083] in: Represents the model's prediction data; Represents real data; is the real data mean; is the index; n is the number of samples.

[0084] See attached Figure 3 , 200 mutants were selected as the first round library in the present embodiment, and Rosetta Flex Calculate the change in their binding free energy Value. Figure 3 From the data, we can see that the distribution of changes in binding free energy for three different types of mutants (single-point mutants, double-point mutants, and triple-point mutants) is displayed. The range of changes in binding free energy for single-point mutants is relatively small, and the changes in binding free energy for double-point mutants are more dispersed due to more mutations. Triple-point mutants have changes at three amino acid sites, and they have the largest range of changes in binding free energy. Because they have the most mutations, these mutants may show the most dispersed distribution in the figure. As the number of mutation sites of the mutants increases (from single point to double point to triple point), their changes in binding free energy may become more dispersed, indicating that increasing mutation sites will increase the diversity of binding energy.

[0085] S42, using the first round of optimal machine learning model in S41 to predict variants not calculated in the library The value of the mutant The predicted mutant The top 50 mutants were ranked from low to high using Rosetta's flex_ APP module calculates mutant The binding free energy changes of the 50 mutants and their corresponding feature vectors are added to the training data, and the machine learning model is retrained using the updated training data. Regression fitting is performed, and the second round of optimal machine learning model is obtained according to the model evaluation index, thereby achieving the purpose of improving the machine learning model.

[0086] S43, using the second round of optimal machine learning model from S42 to predict the remaining uncalculated variants in the library The value of the mutant value, all mutants The values ​​were ranked from low to high, and the top five mutants were selected as potential high-affinity candidate antibodies.

[0087] Table 1: Feature vectors for each amino acid defined using amino acid descriptors

[0088]

[0089] See attached Figure 4 , which are the experimental results of modeling using different machine learning models of AA2+PSSM amino acid descriptors in the embodiments of the present invention. Figure 4 The data shows that the coefficient of determination, mean absolute error and root mean square error of the particle swarm optimization support vector machine model perform best among the three models presented. The coefficient of determination is an important indicator for measuring the prediction performance of the model. The closer its value is to 1, the stronger the prediction ability of the model is. The particle swarm optimization support vector machine model has the largest coefficient of determination among the three models. The mean absolute error is an indicator for measuring the difference between the model prediction results and the actual results. The smaller its value is, the more accurate the prediction results of the model are. The root mean square error is usually used together with the coefficient of determination and the mean absolute error to comprehensively evaluate the performance of the model. The smaller the value is, the more accurate the prediction results of the model are. It can be seen that the particle swarm optimization support vector machine model has high accuracy and stability in the prediction task and has the best performance.

[0090] Example 2

[0091] The potential high-affinity candidate antibodies screened out in Example 1 were cloned, plasmids were constructed for protein expression, and affinity determination experiments were performed to verify the affinity of the expressed proteins, thereby screening out antibodies with high affinity.

[0092] (1) The top five mutants were selected as potential high-affinity candidate antibodies.

[0093] The specific implementation is to select the mutants in the variant library using the improved support vector machine (SVM) machine learning model. The top five mutants after the prediction results were: 1-7-HWK, 1-7-YWK, 1-7-HWY, 1-7-SWK and 1-7-SWQ.

[0094] (2) Construction of target plasmids for 1-7-HWK, 1-7-YWK, 1-7-HWY, 1-7-SWK and 1-7-SWQ mutants by molecular cloning technology

[0095] The specific implementation is to select NdeⅠ and XhoⅠ restriction sites according to the pET22b plasmid map, retain the 6×His Tag at both ends of the vector, and design specific upstream and downstream primers using 1-7-HWK, 1-7-YWK, 1-7-HWY, 1-7-SWK and 1-7-SWQ phages as templates, and amplify the target fragment by polymerase chain reaction (PCR). After the reaction, the amplified product was identified by 1.5% agarose gel electrophoresis (Marker: DM2000). After confirming the size of the corresponding target band, the PCR product was recovered with the Cycle Pure Kit, and eluted with 10% EB elution buffer in the last step. After measuring the product concentration and marking, it was stored at -20 ℃. Refer to the instructions of the plasmid mini-extraction kit to extract the pET28a plasmid, and elute with 10% EB elution buffer in the last step. After determining the plasmid concentration, use 1% agarose gel to identify the product (Marker: 1KbDNA Ladder). The plasmid is marked and stored at -20 °C. Use NdeⅠ and XhoⅠ endonucleases to digest the pET28b plasmid, prepare the enzyme digestion reaction system on ice, centrifuge briefly after preparation, and react at 37 °C on the PCR instrument for 20 min using the advantage of precise temperature control of the PCR instrument. After the enzyme digestion reaction is completed, use 1% agarose gel to identify the results, and use the Gel ExtractionKit kit to recover the enzyme digestion product. After determining the concentration of the recovered product, mark it and store it at -20 °C. Refer to the instructions of the Vazyme recombination kit, with a molar ratio of 1:10 between the vector after enzyme digestion and the target fragment, prepare the recombination reaction system on ice, and express the OmpA signal peptide at the amino terminus at the same time. React at 37 °C on the PCR instrument for 30 min. After the reaction is completed, put the product on ice to cool immediately. Thus, the recombinant target plasmid is obtained.

[0096] (3) Transform the target plasmid into competent cells, expand the culture in the culture medium, and ultrasonically disrupt them to obtain the higher ranked lower The mutant protein crude extract with the value;

[0097] The specific implementation is to transform the target plasmid into BL21 (DE3) competent cells, inoculate them in LB liquid medium for culture, and culture them on a constant temperature shaker at 2000 rpm for about 2 hours. 600When the 1-7-PWR, 1-7-SWR, 1-7-CWR, 1-7-MWR and 1-7-SWQ were obtained, 0.3 mM isopropyl-d-thiogalactoside (IPTG) was added when the 1-7-PWR, 1-7-SWR, 1-7-CWR, 1-7-MWR and 1-7-SWQ were reached. After culturing at 200 rpm for 12 hours under constant temperature, the cells were collected by centrifugation at 8000 rpm for 5 minutes, and then sonicated for 10 minutes using an ultrasonic cell disruptor with an output power of 120 W. After sonication, the cells were centrifuged at 10000 rpm for 30 minutes, and the supernatant was collected to obtain crude protein extracts of 1-7-PWR, 1-7-SWR, 1-7-CWR, 1-7-MWR and 1-7-SWQ.

[0098] (4) Protein expression analysis by SDS-PAGE;

[0099] The specific implementation was to take 10 μL of crude protein extract of 1-7-HWK, 1-7-YWK, 1-7-HWY, 1-7-SWK and 1-7-SWQ, perform SDS-PAGE gel electrophoresis, and perform antibody protein expression analysis. The results showed that 1-7-HWK, 1-7-YWK, 1-7-HWY, 1-7-SWK and 1-7-SWQ were expressed normally.

[0100] (5) Purify using Ni-NTA column (source: Genscript, product number: L00250) and determine the concentration of mutant protein using BCA protein assay kit (source: Beijing Biosen Biotechnology Co., Ltd., product number: C05-02001);

[0101] The specific implementation is to use Ni-NTA column to purify 1-7-HWK, 1-7-YWK, 1-7-HWY, 1-7-SWK and 1-7-SWQ. After purification, the protein concentration is determined by BCA protein kit. The results show that the expression level of 1-7 is 0.12 mg / mL, the expression level of 1-7-HWK is 0.14 mg / mL, the expression level of 1-7-YWK is 0.10 mg / mL, the expression level of 1-7-HWY is 0.15 mg / mL, the expression level of 1-7-SWK is 0.12 mg / mL, and the expression level of 1-7-SWQ is 0.15 mg / mL.

[0102] (6) The indirect competitive ELISA method was used to measure the true affinity of the purified mutant protein to obtain high-affinity antibodies.

[0103] The specific implementation is to measure the binding affinity of purified 1-7-HWK, 1-7-YWK, 1-7-HWY, 1-7-SWK and 1-7-SWQ by indirect competitive ELISA. The results are shown in the attached Figure 5 shown.

[0104] By the attached Figure 5 The data show that the absorbance values ​​of each point of 1-7-HWK, 1-7-YWK, 1-7-HWY, 1-7-SWK and 1-7-SWQ proteins are all above the 1-7 antibody, indicating that the affinity of the mutants is indeed improved compared with the wild-type protein, and the affinity of 1-7-HWK protein is more significantly improved. Finally, after four-parameter Logistic curve fitting, the relative affinity of 1-7-HWK is 115.3 nM, which is 5.04 times higher than that of the wild-type protein. The relative affinity of 1-7-YWK is 128.1 nM, which is 4.54 times higher than that of the wild-type protein. The relative affinity of 1-7-HWY is 126.6 nM, which is 4.59 times higher than that of the wild-type protein. The relative affinity of 1-7-SWQ is 176.3 nM, which is 3.3 times higher than that of the wild-type protein. The relative affinity of 1-7-SWK is 195. nM, which is 2.98 times higher than the wild-type protein, so a high-affinity antibody is obtained by the above method. Obviously, the affinity of the top five mutants screened out has been improved. However, the mutants screened out by computer-aided design through virtual mutation often only improve the affinity of a small number of mutants, and its accuracy is limited; compared with the method of obtaining high-affinity antibodies by computer-aided design, the method of the present invention has higher accuracy and faster speed.

[0105] The present invention is described with reference to methods, algorithms and experimental flow charts according to embodiments of the invention.

[0106] It should be noted that, in this application, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0107] In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0108] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for improving antibody affinity based on molecular dynamics and machine learning, characterized in that: include: S1, obtain the amino acid sequence of the antibody and preprocess to obtain the best model for antigen-antibody docking; S2, based on the best model of antigen-antibody docking, the simulation tool is used to analyze and combine the molecular dynamics calculation method to obtain the best amino acid mutation site of the antibody; The optimal amino acid mutation site of the antibody obtained by combining the molecular dynamics calculation method comprises: S21, based on the best model of antigen-antibody docking, uses simulation tools to perform structural analysis of antigen-antibody docking, identifies potential mutation sites in antibodies, performs single-point saturation mutagenesis on these mutation sites, and obtains antigen-antibody complexes after single-point mutation; S22, calculate the change in binding free energy of the antigen-antibody complex after single-point mutation, and select single-point mutants with a change in binding free energy less than -1 kcal / mol; S23, combining the single-point mutants to generate double-point mutants, calculating the change in binding free energy of the antigen-antibody complex after the double-point mutations, and determining the double-point mutant with the lowest change in binding free energy as the best double-point mutant; S24, combining the best double-point mutant obtained in S23 with the single-point mutant selected in S22 to generate a triple-point mutant, calculating the change in binding free energy of the antigen-antibody complex after the triple-point mutation, and determining the triple-point mutant with the lowest change in binding free energy as the best triple-point mutant, and determining the amino acid mutation site of the best triple-point mutant as the best amino acid mutation site of the antibody; S3, randomly inserting all possible amino acids at the optimal amino acid mutation site of the antibody to generate a variant library; S4, extracting feature vectors of mutants in the variant library as input, using the change in binding free energy as a label, inputting into a pre-trained machine learning model, predicting the change in binding free energy of mutants in the variant library, and screening potential high-affinity candidate antibodies; The pre-trained machine learning model includes: S41, select 200 mutants from the variant library as the first round library, calculate the change in binding free energy of the mutants, define the feature vector of each amino acid by using the amino acid descriptor, define the feature vector of the mutant by connecting the feature vector of the amino acid at each mutation site, use the feature vector of the mutant in the first round library as input, combine the free energy change as a label to train the machine learning model, perform regression fitting on the feature vector of the mutant and the label binding free energy change, and obtain the first round optimal machine learning model according to the model evaluation index; S42, using the first round of optimal machine learning model to predict the binding free energy changes of mutants whose binding free energy changes have not been calculated in the variant library, arranging the predicted binding free energy changes of the mutants from low to high, selecting the top 50 mutants, calculating the binding free energy changes of the mutants, adding the binding free energy changes of these 50 mutants and their corresponding feature vectors to the training data, retraining the machine learning model with the updated training data, performing regression fitting, and obtaining the second round of optimal machine learning model according to the model evaluation index; S43, using the second round of optimal machine learning model to predict the binding free energy changes of the remaining mutants in the variant library whose binding free energy changes have not been calculated, rank the binding free energy changes of all mutants from low to high, and select the top five mutants as potential high-affinity candidate antibodies; The machine learning model is optimized using support vector machine, Gaussian process or random forest combined with particle swarm optimization algorithm.

2. The method for improving antibody affinity based on molecular dynamics and machine learning according to claim 1, characterized in that: The S1 includes: S11, sequencing the antibody by protein sequencing technology to obtain amino acid sequence information of the antibody, wherein the sequence information is composed of sequence information of a plurality of amino acid types connected end to end; S12, inputting the amino acid sequence of the antibody obtained in S11 and the amino acid sequence of the known antigen into the AlphaFold 2 online server to predict the full-length structure of the antibody and the antigen, selecting the structure with the highest ranking based on the score as the best structural information, and determining the position and sequence of the CDRs region of the antibody through the IgBlast tool platform, wherein the best structural information is composed of the three-dimensional coordinates of the atoms of several amino acids connected end to end; S13, inputting the optimal structural information of the antibody and antigen predicted in S12 into the ClusPro 2.0 online server for antigen-antibody molecular docking, and selecting the docking model with the highest ranking as the optimal model for antigen-antibody docking.

3. The method for improving antibody affinity based on molecular dynamics and machine learning according to claim 1, characterized in that: The simulation tool is any one or a combination of InterProSurf, PyMol or Rosetta protein simulation tools; the change in binding free energy is calculated using the following formula; （1） in: represents the gas constant; represents absolute temperature; represents the dissociation constant of the complex; represents the binding free energy; （2） in, represents the binding free energy of the antigen-antibody complex after the antibody is mutated; represents the binding free energy of the original antigen-antibody complex; It represents the change in binding free energy. A positive value indicates that the binding free energy of the complex after mutation increases and the affinity decreases; a negative value indicates that the binding free energy of the complex after mutation decreases and the affinity increases.

4. The method for improving antibody affinity based on molecular dynamics and machine learning according to claim 1, characterized in that: The model evaluation indicators include absolute coefficient , mean absolute error and RMS error ; The absolute coefficient Calculated by formula (3): （3） The mean absolute error Calculated by formula (4): （4） The root mean square error Calculated by formula (5): （5） in: Represents the model's prediction data; Represents real data; is the real data mean; is the index; n is the number of samples.

5. The method for improving antibody affinity based on molecular dynamics and machine learning according to claim 1, characterized in that: The method also includes cloning the screened potential high-affinity candidate antibodies, constructing plasmid transfection for protein expression, verifying the affinity of the expressed protein by affinity determination experiments, and screening antibodies with high affinity.

6. The system used in the method for improving antibody affinity based on molecular dynamics and machine learning according to any one of claims 1 to 5, characterized in that: include: Data preprocessing and model building module, used to obtain the amino acid sequence of the antibody and preprocess to obtain the best model for antigen-antibody docking; The mutation site identification and variant library generation module uses simulation tools and molecular dynamics calculation methods to identify the best amino acid mutation sites of antibodies and generate variant libraries based on the optimal model of antigen-antibody docking; The machine learning model training and prediction module is used to predict the change in binding free energy of mutants in the variant library and screen potential high-affinity candidate antibodies.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method for improving antibody affinity based on molecular dynamics and machine learning as described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for improving antibody affinity based on molecular dynamics and machine learning as described in any one of claims 1 to 5 are implemented.

Images