Method for generating antibody library

Abstract

The invention provides an antibody library generation method which comprises the following steps: initializing an antibody library to construct an initial antibody library, and generating a diversity weight matrix according to the initial antibody library; constructing a self-adaptive generation model according to the initial antibody library and the diversity weight matrix, and generating an optimized antibody library by the self-adaptive generation model by optimizing combination region mutation control and sequence recombination of antibodies; constructing a binding feature matrix of key binding sites of the target antigen according to the antibody library; performing specific screening on the antibody library according to the binding characteristic matrix to generate a high-affinity antibody subset; and constructing a distributed closed-loop optimization system according to the high-affinity antibody subset, and generating a final antibody library. Through three innovation points of adaptive diversity control, feature co-occurrence screening and distributed parallel generation, the bottleneck of the existing antibody library generation technology in the aspects of diversity, binding affinity and screening efficiency is successfully overcome.

Description

Technical Field

[0001] The present invention belongs to the field of antibody libraries, and in particular relates to a method for generating an antibody library. Background Art

[0002] In the field of biopharmaceuticals, the development of antibody drugs relies on efficient antibody library generation and screening technology to quickly find antibody molecules with high affinity and specificity for specific antigens. Traditional antibody library generation mainly relies on phage display, in vitro recombination and other technologies. These methods usually construct a diverse antibody library based on random mutation or sequence recombination, and then screen out antibodies with potential affinity through high-throughput screening methods. However, these methods still have many limitations in terms of the diversity, binding specificity, affinity and screening efficiency of the antibody library.

[0003] First, the traditional random mutation or sequence recombination generation method easily leads to the sequence distribution of the antibody library being too concentrated, that is, the structural variability of the antibody sequence is limited, and it is difficult to cover enough binding site diversity. This homogenized antibody library usually requires a lot of screening to find antibodies with high affinity and specificity, which greatly increases the screening time and cost. Secondly, since antibody-antigen binding is based on specific spatial conformation and epitope matching, conventional generation methods lack refined binding site regulation means, resulting in only a small number of antibodies in the generated antibody library binding well to the target antigen. In addition, traditional screening technology is difficult to efficiently identify and retain high-affinity antibodies when facing large-scale antibody libraries. Usually, the screening process is lengthy and inefficient, especially when the number of antibodies is huge, the screening is time-consuming and labor-intensive. Although auxiliary antibody design and screening methods based on machine learning and deep learning have begun to emerge in recent years, most models still rely on fixed training sets and features, and it is difficult to dynamically adapt to the binding requirements of different antigens when generating antibody libraries, and thus fail to completely solve the problems of low generation and screening efficiency and insufficient antibody diversity.

[0004] Therefore, the current antibody library generation technology still has obvious needs for improvement in terms of diversity, binding affinity, screening efficiency, etc. In order to overcome these technical bottlenecks, an efficient, flexible and adaptive antibody library generation method is urgently needed to achieve rapid and accurate high-affinity antibodies while improving the diversity of the antibody library. Summary of the invention

[0005] The purpose of the present invention is to design a method for generating an antibody library, and to achieve efficient, high-diversity and high-affinity antibody library generation through multi-step adaptive and parallel optimization strategies.

[0006] In order to achieve the above object, the present invention provides a method for generating an antibody library, the method comprising the following steps:

[0007] Step 1: Initialize the antibody library to construct the initial antibody library S init , and according to the initial antibody library S init Generate a diversity weight matrix W, wherein the diversity weight matrix W controls the mutation frequency of each binding site region during antibody generation, and each element w of the diversity weight matrix W i,j The mutation priority of amino acid j in binding region i, i.e., the mutation weight, is expressed as follows:

[0008] w i,j =α·f i,j +β·(1-v i [j])

[0009] Among them, f i,j represents the frequency of occurrence of amino acid j in binding region i, α and β are balance weight parameters, and v i [j] represents the variability distribution of amino acid j in binding region i;

[0010] Step 2: Based on the initial antibody library S init The adaptive generation model is constructed by optimizing the mutation control and sequence recombination of the binding region of the antibody and the diversity weight matrix W to generate an optimized antibody library S; wherein the adaptive generation model is constructed as follows:

[0011] In the initial antibody library S init For each antibody sequence s i The binding area r i Perform adaptive mutation initialization to obtain the binding region sequence set {r i '};

[0012] From the binding region {r i '} Randomly select two binding region sequences r a ' and r b 'Perform sequence crossover recombination to generate new binding regions r a,b ', and repeat the process until a new antibody sequence library S is generated new ;

[0013] Introducing a diversity regularization term Ω(S) into the adaptive generation model, by controlling the structural differences between sequences, preventing the generated antibody library from being homogenized due to the recombination process, and generating the final antibody library S;

[0014] Wherein, the initial antibody library S init For each antibody sequence s i The binding area r i Perform adaptive mutation initialization, specifically including the following steps:

[0015] Set the mutation probability P of amino acid position j in each binding region i,j , used to control the mutation frequency and variability of the binding site, is expressed as follows:

[0016] P i,j =γ·w i,j ·(1+λ·dist(r i [j],r avg ))

[0017] Among them, γ is the global mutation intensity coefficient, which is used to control the mutation intensity; λ is the balance coefficient, which is used to adjust the mutation contribution to the average distance; dist(r i [j],r avg ) indicates that amino acid j is in r i The average sequence of the binding region r avg The distance is used to quantify the difference between the current amino acid and the average sequence;

[0018] According to the mutation probability P i,j For each binding region r i Perform adaptive mutation operations, including:

[0019] For the binding area r i For each amino acid position j in , generate a random number u∈[0,1];

[0020] If u≤P i,j , then perform mutation operation on amino acid position j and change r i [j] mutated to a different amino acid (with v i The amino acid distribution in is used as a reference);

[0021] If u>P i,j , the original amino acid r is retained i [j] unchanged;

[0022] The mutated binding region is denoted as r i ′, forming a new binding region sequence set {r i '};

[0023] Step 3, constructing a binding feature matrix F of key binding sites of the target antigen based on the antibody library S;

[0024] Step 4: specifically screening the antibody library S according to the binding feature matrix F to generate a high-affinity antibody subset S′, wherein the high-affinity antibody subset S′ contains antibody sequences that meet the requirements of the binding feature matrix F;

[0025] Step 5: Construct a distributed closed-loop optimization system based on the high-affinity antibody subset S′ to generate the final antibody library S* .

[0026] Further, according to the initial antibody library S init Generate the diversity weight matrix W, including:

[0027] Identification of initial antibody library S init Each sequence s i , the sequence s i Key amino acid fragments in the binding site region i The set R of all antibody binding sites, R = {r1, r2, ..., r n};

[0028] Construct a diversity feature space and define the variability of each binding region as a vector v i , and the variability of each binding region is the vector v i represents the probability of mutation of different amino acids at the binding site, as follows:

[0029]

[0030] Where count(X) represents the number of amino acids X in the binding site region r. i The frequency of occurrence, L is the total length of the binding region;

[0031] Combined with the diversity feature vector v i And the initial antibody library S init The statistical characteristics of different binding regions in the matrix are used to generate the diversity weight matrix W.

[0032] Furthermore, the diversity weight matrix W is normalized before output so that it satisfies ∑ j w i,j =1.

[0033] Further, the binding region {r i '} Randomly select two binding region sequences r a ' and r b The recombination process for sequence crossover recombination follows the following rules:

[0034] The recombination probability of each binding region is dynamically regulated according to its mutation frequency and structural feature distribution;

[0035] Ensure that the sequence retains the characteristics of the key binding site during crossover recombination to avoid loss of affinity potential due to mutation;

[0036] The diversity regularization term Ω(S) is expressed as follows:

[0037]

[0038] Among them, sim(s i ,s j ) is the sequence s i and j , δ is the equilibrium parameter that controls the similarity regularization term. Through the regularization term Ω(S), the mutation and recombination frequencies are dynamically adjusted in the antibody generation model to ensure sufficient diversity between sequences to increase the adaptability of the antibody library to a wide range of antigen binding sites.

[0039] Furthermore, the step 3 specifically includes:

[0040] Obtain the three-dimensional structural data of the target antigen, extract the surface information of the antigen, and use the surface information of the antigen to form a preliminary feature representation of the target antigen. j ; Wherein, the surface information of the antigen is the surface site representation of the surface area of ​​the target antigen, and the local charge density, polarity value and hydrophobicity coefficient are extracted from the surface site representation to form a preliminary feature representation q j ,q j =(φ j ,ψ j ,ρ j ), φ j Indicates site t j The local charge density of j Indicates site t j The polarity value of j Indicates site t j The hydrophobicity coefficient;

[0041] The preliminary feature representation q j The target antigen surface feature vector set {q j}, and the target antigen surface feature vector set {q j} Perform cluster analysis and calculate the center point b of the target antigen i and the corresponding eigenvector q i is a representative of the combined area;

[0042] The feature vector q i Integrate into the combined feature matrix F, expressed as follows:

[0043]

[0044] Where φ1 represents the local charge density at site t1, φ2 represents the local charge density at site t2, and φ m Indicates site t m The local charge density of site t1 is ψ1, ψ2 is the polarity value of site t2, and ψ m Indicates site t mThe polarity value of ρ1 represents the hydrophobicity coefficient of site t1, ρ2 represents the hydrophobicity coefficient of site t2, and ρ m Indicates site t m The hydrophobicity coefficient;

[0045] Introducing the regularization term Ω(F) into the binding feature matrix F improves the matching accuracy of the binding feature matrix F, thereby ensuring that the antibody library S can adapt to the changes in the specific binding region during the screening process;

[0046] By combining the feature matrix F and the feature distribution of the antibody library S, the fitness of the antibody and the target antigen binding site is calculated, which is expressed as follows:

[0047]

[0048] Among them, score(s k ,b i ) indicates antibody s k Binding site b i The fitness score, f i,j and k,j Represents binding site b i and antibodies k The match score on a specific feature.

[0049] Furthermore, the regularization term Ω(F) is expressed as follows:

[0050]

[0051] Among them, dist(q i ,q j ) is the binding site feature q i and q j is the Euclidean distance between the feature vectors, τ is the adjustment coefficient, and m represents the total number of sites.

[0052] Furthermore, the specific screening of the antibody library S according to the key binding site feature matrix F to generate an antibody subset S′ with high affinity specifically includes:

[0053] According to the structural characteristics of the combined feature matrix F, each antibody sequence s in the antibody library S k Calculate its relationship with each binding site b i The specific match degree, that is, the match score of each antibody sequence k,i ; Among them, the matching score of each antibody sequence is match k,i The lowest, indicating that the antibody sequence s k Binding site b i The characteristics are closest to those of the antigen, meeting the binding requirements of the antigen;

[0054] According to each antibody sequence s k Match score match k,i and the overall affinity of all binding sites to construct each antibody sequence s k The screening score score k , and introduce the regularization term Ω(s k ,F) equilibrium antibody affinity distribution at different binding sites;

[0055] According to the filter score score k The antibody library S is sorted, and the top N antibody sequences with the lowest scores are selected to form a high-affinity antibody subset S′.

[0056] Furthermore, the matching score match k,i It is expressed as follows:

[0057]

[0058] Among them, f i,j Represents the binding site b in the binding feature matrix F i The jth characteristic value of i , polarity ψ i or hydrophobicity ρ i ;s k,j Represents antibodies k The jth characteristic value in the combined area; ω j is the feature importance weight, which controls the contribution of each feature to the matching score;

[0059] The screening score k , calculated as follows:

[0060]

[0061] Among them, α is the regularization term weight, which balances the influence of the regularization term and the matching score;

[0062] The regularization term Ω(s k ,F), expressed as follows:

[0063]

[0064] Among them, dist(q i ,q j ) is the binding site characteristic q i and q j is the Euclidean distance between the feature vectors, and τ is the adjustment coefficient.

[0065] Furthermore, the distributed closed-loop optimization system is constructed based on the high-affinity antibody subset S′ to generate the final antibody library S * , including:

[0066] Assign the high-affinity antibody subset S′ to multiple nodes, that is, the antibody subset of each node p is S′ p , where p = 1, 2, …, P, where P is the total number of nodes;

[0067] Perform a specificity score on each node p k and the diversity regularization term Ω(s k ,F), ensure that the antibody subset reaches the local optimum within the node;

[0068] After completing local optimization at each node p, the specificity of each antibody is scored k and its binding site characteristicsq k The data are transmitted to the master node, which evaluates the binding diversity of the entire antibody library based on the feedback information and calculates the global diversity score diversity(S′). The master node updates the key parameters in the binding feature matrix F and the diversity weight matrix W based on the scores fed back by each node.

[0069] After receiving the updated key parameters, each computing node recalculates the specificity score of each antibody in each antibody sequence. k , and for the antibody subset S′ p Rescreen the antibodies in the test set and retain the top N antibodies with the lowest scores;

[0070] After multiple rounds of closed-loop iterations, the antibody subsets of each node will tend to the global optimum, that is, a more balanced state in terms of specificity and diversity. Finally, the antibody subsets of each node will be merged to form the final antibody library S * .

[0071] Furthermore, the specificity score k It is expressed as follows:

[0072]

[0073] Among them, match k,i Antibodies k Binding site b i The specific matching score describes the adaptability of the antibody at different binding sites; Ω(s k ,F) is the regularization term, which mainly controls the characteristic smoothness of the binding region to avoid excessive concentration on a few binding sites; α is the regularization coefficient, which is used to regulate the weight of the regularization term in the overall score;

[0074] The diversity score (S′) is expressed as follows:

[0075]

[0076] Among them, dist(q p,i ,q p,j ) represents the antibody s on node p k and l The Euclidean distance on the binding feature, τ is a balance parameter, which controls the effect of the feature distance on the diversity score; the diversity score diversity(S′) is used to measure the coverage of the antibody subset S′ on the binding region.

[0077] The beneficial technical effects of the present invention are at least as follows:

[0078] First, in order to solve the problem of homogeneity of antibody libraries, the present invention constructs an adaptive generation model based on a diversity weight matrix, which realizes the diversity optimization of antibody binding regions by dynamically adjusting the mutation and structural distribution of antibody sequences. This mechanism automatically optimizes the binding region sequences in the antibody library according to the structural characteristics of the antigen, so that the antibody library has a high degree of structural diversity, effectively solving the problem of high homogeneity of antibody sequences in traditional generation methods.

[0079] Secondly, in order to improve the binding efficiency and affinity of the antibody library, the present invention proposes a co-occurrence screening model based on the characteristics of the target antigen binding site. By extracting the characteristics of the antigen binding site and generating a binding feature matrix, antibody sequences with a high degree of match with the co-occurrence characteristics of the antigen binding site are screened out to ensure that the screened antibodies have high affinity and specificity. This screening mechanism effectively makes up for the lack of binding region targeting in traditional random generation and screening methods, and greatly improves the binding efficiency and screening accuracy of the antibody library.

[0080] Finally, in order to further improve the efficiency of antibody library generation, the present invention introduces a distributed computing architecture, parallelizes the antibody generation and screening tasks, and constructs a real-time feedback mechanism between the master node and the distributed nodes. The master node is responsible for collecting the generation results of each node and optimizing the diversity weights and binding site characteristics, forming a closed-loop feedback system, and continuously improving the efficiency of generation and screening. This distributed architecture significantly speeds up the generation of antibody libraries, and is particularly suitable for the high-throughput screening requirements of large-scale antibody libraries.

[0081] In summary, the present invention successfully overcomes the bottlenecks of existing antibody library generation technology in terms of diversity, binding affinity and screening efficiency through three major innovations: adaptive diversity control, feature co-occurrence screening and distributed parallel generation, and significantly improves the adaptability, generation efficiency and screening effect of antibody libraries. This patented solution can not only quickly generate high-quality antibody libraries to meet complex and diverse antigen requirements, but also significantly reduce the time and cost of antibody screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0083] Figure 1 The present invention is a flow chart of a method for generating an antibody library. DETAILED DESCRIPTION

[0084] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0085] In one or more embodiments, Figure 1 As shown, a method for generating an antibody library of the present invention is disclosed, and the method comprises steps 1-5:

[0086] Step 1: Initialize the antibody library to construct the initial antibody library S init , and according to the initial antibody library S init Generate the diversity weight matrix W.

[0087] Specifically, assuming that the existing antibody sequence library is S init , where each antibody sequence is represented by s i , containing specific amino acid sequence and structural information. init Perform redundancy processing to remove sequences with severe homogeneity. init Divided into several diversity groups, denoted as C1, C2, …, C k , to ensure that each group represents a structural or binding site difference.

[0088] For each sequence s i , the present invention identifies the key amino acid fragments in the binding site region. Assume that the length of the binding region of each sequence is L, denoted by r i Each r i The key feature sites of the binding region are determined as follows:

[0089] R={r1,r2,…,r n}

[0090] Where R represents the set of all antibody binding sites. For each binding site r i The present invention analyzes the amino acid type and physicochemical properties (such as polarity, hydrophobicity, etc.) to form a characteristic distribution matrix M of the binding region.char .

[0091] Furthermore, in order to quantify the diversity of the antibody library, the present invention constructs a diversity feature space and defines the variability of each binding region as a vector v i Each vector element represents the probability of variation of different amino acids at the binding site. The variability can be calculated by the following formula:

[0092]

[0093] Where count(X) represents the number of amino acids X in the binding site region r. i The frequency of occurrence of L is the total length of the binding region. i After normalization, the diversity feature vector v of each combined region is obtained i , in order to analyze the variability of the overall antibody repertoire.

[0094] Furthermore, combined with the diversity feature vector v i As well as the statistical characteristics of different binding regions in the initial antibody library, a diversity weight matrix W is generated. W is used to control the mutation frequency of each binding site region during antibody generation. Specifically, each element w in W i,j represents the mutation priority of amino acid j in binding region i and is defined as follows:

[0095] w i,j =α·f i,j +β·(1-v i [j])

[0096] Among them, f i,j represents the frequency of occurrence of amino acid j in binding region i, α and β are balance weight parameters, and v i [j] represents the variability distribution of amino acid j in binding region i. This matrix W will be used as the input of the subsequent adaptive generation model to ensure the structural diversity of the binding site region.

[0097] Finally, the diversity weight matrix W is normalized to satisfy ∑ j w i,j = 1. In this way, W can both reflect the mutation requirements of the binding region and avoid weight bias in mutation control.

[0098] After the above steps, the initial antibody library S is output. init , which is used for subsequent adaptive generation; the diversity weight matrix W provides guidance for mutation regulation in antibody generation to ensure that the subsequently generated antibody library has a higher structural diversity of binding sites.

[0099] Step 2: Based on the initial antibody library S initThe adaptive generation model is constructed by the diversity weight matrix W, and the adaptive generation model generates an optimized antibody library S by optimizing the mutation control and sequence recombination of the binding region of the antibody.

[0100] Specifically, first, the present invention is in the initial antibody library S init For each antibody sequence s i The binding area r i Perform adaptive mutation initialization. Combined with the diversity weight matrix W, the present invention sets the mutation probability P of the amino acid position j in each binding region. i,j , to control the mutation frequency and variability of the binding site. Specifically, define the mutation probability P i,j for:

[0101] P i,j =γ·w i,j ·(1+λ·dist(r i [j],r avg ))

[0102] Among them, γ is the global mutation intensity coefficient, which is used to control the mutation intensity (typical value is 0.01 to 0.05); w i,j is the mutation weight of amino acid j in binding region i (taken from W), reflecting the importance of mutation at each position; λ is the balance coefficient, which is used to adjust the mutation contribution to the average distance; dist(r i [j],r avg ) indicates that amino acid j is in r i The average sequence of the binding region r avg The distance between i [j],r avg ) is calculated by taking into account the physicochemical properties of amino acids (such as hydrophobicity and polarity) to quantify the difference between the current amino acid and the average sequence. This mutation probability model introduces a specific balance term λ·dist, which increases the diversity of the binding region and enhances the coverage of different antigen binding sites in the antibody library by controlling the mutation amplitude from the average distance.

[0103] Furthermore, according to the mutation probability P set in the previous step i,j , for each binding region r i Perform adaptive mutation operation. The specific steps are as follows:

[0104] For the binding area r i For each amino acid position j in , generate a random number u∈[0,1];

[0105] If u≤P i,j , then perform mutation operation on amino acid position j and change r i [j] mutated to a different amino acid (with vi The amino acid distribution in is used as a reference); if u>P i,j , the original amino acid r is retained i [j]Unchanged.

[0106] The mutated binding region is denoted as r i '. After this adaptive mutation operation, a new binding region sequence set {r i ′}, whose amino acid distribution is more in line with the requirements of the diversity weight matrix W, and is helpful to generate diverse binding regions in the antibody library.

[0107] Further, after obtaining the adaptive mutation binding region {r i '}, the diversity of the antibody library is further enhanced by sequence recombination. i '} Randomly select two binding region sequences r a ' and r b 'Perform sequence crossover recombination to generate new binding regions r a,b ', and repeat the process until a new antibody sequence library S is generated new For each antibody sequence s after sequence recombination a,b , the new sequence of its binding region r a,b 'Exhibits further diversity and forms the basis of antibody library S.

[0108] The reorganization process follows the following rules:

[0109] The recombination probability of each binding region is dynamically regulated according to its mutation frequency and structural feature distribution;

[0110] Ensure that the sequence retains the characteristics of key binding sites during crossover recombination to avoid loss of affinity potential due to mutation.

[0111] Furthermore, in order to further control the diversity of the antibody library S, the diversity regularization term Ω(S) is introduced to optimize the antibody structure diversity in the generation model. The regularization term Ω(S) prevents the generated antibody library from being homogenized due to the recombination process by controlling the structural differences between sequences:

[0112]

[0113] Among them, sim(s i ,s j ) is the sequence s i and jThe similarity metric (e.g., based on the amino acid features of the binding region) is used, and δ is the equilibrium parameter that controls the similarity regularization term. Through the regularization term Ω(S), the mutation and recombination frequencies are dynamically adjusted in the antibody generation model to ensure that sufficient diversity is maintained between sequences to increase the adaptability of the antibody library to a wide range of antigen binding sites.

[0114] Finally, the antibody sequence set S after adaptive mutation and sequence recombination and regular optimization is new The output antibody library of this step is denoted as S. This antibody library contains richer binding region diversity and structural combinations, providing a more comprehensive basis for subsequent antibody specificity screening.

[0115] The output of this step is the generated antibody library S, in which each antibody sequence has unique structural characteristics and diversity in the binding region. This antibody library S has a wider coverage of the binding region, which helps to improve the affinity and specificity of the screening antibodies in the subsequent steps.

[0116] Step 3: Construct a binding feature matrix F of the key binding sites of the target antigen based on the antibody library S.

[0117] Specifically, first, the present invention obtains the three-dimensional structural data of the target antigen, for example, through the structural data file in the protein data bank (PDB), and extracts the surface information of the antigen. Assume that the surface area set of the target antigen is T, and each surface site is represented by t j For each t j The present invention calculates the following characteristics to form the feature vector q j , including: φ j :Locus t j The local charge density of j :Locus t j The polarity value of j :Locus t j The hydrophobicity coefficient of j The preliminary feature representation of is:

[0118] q j =(φ j ,ψ j ,ρ j )

[0119] Among them, q j Indicates t j Preliminary features of the binding region. The feature vector constructed in this way will provide a basis for subsequent binding site identification, enabling it to cover the distribution characteristics of the antigen surface binding potential.

[0120] Furthermore, the target antigen surface feature vector set {q j} Perform cluster analysis to identify regions with higher binding potential. Using density-based clustering methods (such as DBSCAN), the present invention combines feature vectors of adjacent binding sites into potential binding regions. For each cluster, let its center point be b i , and its eigenvector q i Specifically, all site features in the binding site region are weighted averaged to generate the binding center point q i The final eigenvalue of :

[0121]

[0122] Among them, N i Indicates the binding region b i The neighborhood site set of |N i | is the number of sites in the region. Thus, q i The feature vector of the binding center contains the comprehensive information of the charge, polarity and hydrophobicity distribution of the binding area.

[0123] Furthermore, the characteristic vector q of the center of each binding site is i Integrate into a binding feature matrix F, which is used as the matching standard for subsequent antibody screening. Specifically, construct a binding feature matrix F, where each row represents a key feature of a binding site, in the following format:

[0124]

[0125] Among them, φ i ,ψ i ,ρ i Represent the charge density, polarity and hydrophobicity characteristics of binding site i respectively. Matrix F will become the characteristic benchmark when matching antibodies and antigens in antibody library S, ensuring that the charge, polarity and hydrophobicity characteristics of the binding site meet the antigen requirements.

[0126] Furthermore, in order to improve the matching accuracy of the feature matrix F, the regularization term Ω(F) is introduced to ensure that the antibody library S can adapt to subtle changes in specific binding regions during the screening process. The regularization term Ω(F) is defined as:

[0127]

[0128] Among them, dist(q i ,q j ) is the binding site characteristic q i and q j The regularization term is the Euclidean distance between the feature vectors, and τ is the adjustment coefficient. This regularization term avoids the binding sites from being too dense in terms of structural features by regulating the distance, thus preventing mismatching or feature redundancy in the screening model.

[0129] Furthermore, by combining the feature matrix F and the feature distribution of the antibody library S, the present invention calculates the fitness of the antibody and the target antigen binding site. k Binding site b i The fitness score is:

[0130]

[0131] Among them, f i,j and k,j Represents binding site b i and antibodies k The matching score on a specific feature (such as charge, polarity or hydrophobicity). By calculating the matching score, the present invention can score the binding properties of antibodies in the antibody library S to ensure that only antibody sequences that meet the requirements of the feature matrix F are screened out.

[0132] Finally, the constructed binding feature matrix F is used as the output of this step and provided to the subsequent antibody screening model. The matrix F structurally describes the characteristics of the antigen binding site, so that the subsequent screening process can accurately evaluate the binding ability of the antibody and ensure that the antibody sequence is highly matched with the antigen binding region.

[0133] Step 4: Perform specific screening on the antibody library S according to the binding feature matrix F to generate a high-affinity antibody subset S′, wherein the high-affinity antibody subset S′ contains antibody sequences that meet the requirements of the binding feature matrix F.

[0134] Specifically, the present invention combines the structural characteristics of the feature matrix F to identify each antibody sequence s in the antibody library S. k Calculate its relationship with each binding site b i Specifically, in order to ensure the exact match of antibody sequences in terms of charge, polarity and hydrophobicity, the present invention defines the antibody sequence s k Binding site b i Match score match k,i :

[0135]

[0136] Among them, f i,j Represents the binding site b in the binding feature matrix F i The jth characteristic value (charge φ i , polarity ψ i or hydrophobicity ρ i );s k,j Represents antibodies k The jth characteristic value in the combined area; ω j It is the feature importance weight, which controls the contribution of each feature to the matching score.

[0137] Among them, the matching score is k,i The lower the antibody sequence, the k Binding site b i The closer the features of the antibody are to the binding site, the better it meets the binding requirements of the antigen. This scoring strategy ensures that the antibody is well matched to the binding site in all characteristics.

[0138] Furthermore, in the preliminary calculation of the score match k,i Finally, in order to balance the overall affinity of the antibody sequence at all binding sites, the present invention defines a global affinity score k , and introduce the regularization term Ω(s k ,F) to balance the affinity distribution of antibodies at different binding sites. k The definition is as follows:

[0139]

[0140] Among them, α is the regularization term weight, which balances the influence of the regularization term and the matching score.

[0141] Among them, the regularization term Ω(s k ,F) is used to prevent uneven performance of antibodies at different binding sites and ensure the stability of overall affinity, and is defined as follows:

[0142]

[0143] Among them, dist(q i ,q j ) is the binding site characteristic q i and q j The Euclidean distance between the feature vectors, τ is the adjustment coefficient. The regular term Ω(s k ,F) By controlling the affinity changes between binding sites, it is ensured that the screening results are not only affinity for a single binding site, but also stable in different binding regions.

[0144] Further, after obtaining each antibody s k The screening score score k Finally, the antibody library S is sorted based on this score, and the top N antibody sequences with the lowest scores are selected to form the high-affinity antibody subset S'. This screening method ensures that the antibodies in S' have excellent performance in affinity, matching specificity, and binding region smoothness.

[0145] Furthermore, in order to further optimize the antibody subset S′, the present invention adopts a step-by-step feedback optimization strategy. Specifically, by adjusting the weight ω jand regularization coefficient α, optimize the contribution of different characteristics to affinity screening to obtain screening results with higher specificity. The score is recalculated after each round of feedback optimization. k And S′ is updated to ensure that the affinity and specificity of the antibody subset are continuously improved in each round of screening.

[0146] This step outputs a high-affinity antibody subset S', which contains antibody sequences that meet the requirements of the binding feature matrix F. This antibody subset achieves a high degree of adaptation to the antigen binding region in multiple characteristics, providing a basis for subsequent antibody optimization.

[0147] Step 5: Construct a distributed closed-loop optimization system based on the high-affinity antibody subset S′ to generate the final antibody library S * .

[0148] Specifically, the high-affinity antibody subset S′ is distributed to multiple computing nodes to support large-scale parallel processing. Let the antibody subset of each computing node p be S′ p (p = 1, 2, ..., P, where P is the total number of nodes.) In a distributed system, each node will independently perform binding site specificity optimization and antibody sequence diversity optimization.

[0149] Wherein, node (p): is a distributed computing unit that performs local optimization calculations. Each node independently processes the assigned antibody subset S' p , calculate the specificity score k and the diversity regularization term Ω(s k ,F), and screen and optimize antibody subsets.

[0150] Master node: It is the central unit responsible for global information integration and parameter update. The master node collects the optimization results fed back by each computing node, evaluates the overall binding diversity (S') of the antibody library, updates key parameters (combining feature matrix F and diversity weight matrix W), and sends the updated parameters back to the computing node for continued iterative optimization.

[0151] Furthermore, in order to adapt to distributed processing, each node loads the shared binding feature matrix F and diversity weight matrix W, and calculates the importance of each binding region based on the characteristic distribution of F and the binding weight in W. p Mutation and optimization are performed through these two matrices to improve the binding specificity and diversity of the antibody library.

[0152] Furthermore, each node performs specificity scoring and combined diversity regularization to ensure that the antibody subset reaches the local optimum within the node. k ∈S′ p, calculate its optimization score under the constraint of the feature matrix F k , including the specific matching score and the diversity regularization term. The formula is as follows:

[0153]

[0154] Among them, match k,i Antibodies k Binding site b i The specific matching score describes the adaptability of the antibody at different binding sites; Ω(s k ,F) is the regularization term, which mainly controls the characteristic smoothness of the binding region to avoid excessive concentration on a few binding sites. α is the regularization coefficient, which is used to adjust the weight of the regularization term in the overall score.

[0155] Furthermore, after each node completes local optimization, the comprehensive score of each antibody is calculated. k and its binding site characteristicsq k The master node evaluates the binding diversity of the entire antibody library based on the feedback information and calculates the global diversity score (diversity(S′)). The calculation of the diversity score is intended to measure whether the antibody subset S′ is evenly distributed in the characteristics of each binding site to ensure that the antibody subset can cover the different binding requirements of the antigen. The global diversity score is defined as follows:

[0156]

[0157] Among them, dist(q p,i ,q p,j ) represents the antibody s on node p k and l The Euclidean distance on the binding feature, τ is a balance parameter that controls the effect of feature distance on the diversity score. The diversity score is used to measure the coverage of the antibody subset S' on the binding region.

[0158] The master node updates the key parameters in the binding feature matrix F and the diversity weight matrix W according to the scores fed back by each node, such as the binding site importance weight ω j and the regularization coefficient α, and sends the updated parameters to each distributed node to continue the optimization.

[0159] Furthermore, the master node globally optimizes and updates the feedback parameters, and feeds back the updated F and W to each distributed node, thus forming a closed-loop iteration. After receiving the updated parameters, each node recalculates the optimization score of each antibody sequence. k , and for the subset S′ pThe antibodies in the pool were rescreened, the antibodies with higher scores were removed, and the top N antibodies with the lowest scores were retained.

[0160] After multiple rounds of closed-loop iterations, the antibody subsets of each node will tend to the global optimum, that is, a more balanced state in terms of specificity and diversity. Finally, the antibody subsets of each node are merged to form the final antibody library S * .

[0161] Furthermore, after completing multiple rounds of closed-loop optimization, the master node merges the antibody subsets optimized by each node to form the final antibody library S * The antibody sequences in this library meet the patent requirements in terms of binding diversity, affinity and specificity, can cope with complex antigen structures, and meet the patent's requirements for high efficiency and specificity in the generation of antibody libraries.

[0162] The final output antibody library S * It contains antibody sequences screened in multiple rounds of closed-loop optimization and has excellent binding area coverage, binding affinity and diversity.

[0163] In summary, the present invention successfully overcomes the bottlenecks of existing antibody library generation technology in terms of diversity, binding affinity and screening efficiency through three major innovations: adaptive diversity control, feature co-occurrence screening and distributed parallel generation, and significantly improves the adaptability, generation efficiency and screening effect of antibody libraries. This patented solution can not only quickly generate high-quality antibody libraries to meet complex and diverse antigen requirements, but also significantly reduce the time and cost of antibody screening.

[0164] The above disclosures are only some preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention are still within the scope of the invention.

Claims

1. A method for generating an antibody library, characterized in that: The method comprises the following steps: Step 1: Initialize the antibody library to construct the initial antibody library S init , and according to the initial antibody library S init Generate a diversity weight matrix W, wherein the diversity weight matrix W controls the mutation frequency of each binding site region during antibody generation, and each element w of the diversity weight matrix W i,j The mutation priority of amino acid j in binding region i, i.e., the mutation weight, is expressed as follows: w i,j =α·f i,j +β·(1-v i [j]) Among them, f i,j represents the frequency of occurrence of amino acid j in binding region i, α and β are balance weight parameters, and v i [j] represents the variability distribution of amino acid j in binding region i; Step 2: Based on the initial antibody library S init The adaptive generation model is constructed by the diversity weight matrix W, and the adaptive generation model generates an optimized antibody library S by optimizing the mutation control and sequence recombination of the binding region of the antibody; wherein the adaptive generation model is constructed as follows: In the initial antibody library S init For each antibody sequence s i The binding area r i Perform adaptive mutation initialization to obtain the binding region sequence set {r i '}; From the binding region {r i '} Randomly select two binding region sequences r a ' and r b 'Perform sequence crossover recombination to generate new binding regions r a,b ', and repeat the process until a new antibody sequence library S is generated new ; Introducing a diversity regularization term Ω(S) into the adaptive generation model, by controlling the structural differences between sequences, preventing the generated antibody library from being homogenized due to the recombination process, and generating the final antibody library S; Wherein, the initial antibody library S init For each antibody sequence s i The binding area r i Perform adaptive mutation initialization, specifically including the following steps: Set the mutation probability P of amino acid position j in each binding region i,j , used to control the mutation frequency and variability of the binding site, is expressed as follows: P i,j =γ·w i,j ·(1+λ·dist(r i [j],r avg )) Among them, γ is the global mutation intensity coefficient, which is used to control the mutation intensity; λ is the balance coefficient, which is used to adjust the mutation contribution to the average distance; dist(r i [j],r avg ) indicates that amino acid j is in r i The average sequence of the binding region r avg The distance is used to quantify the difference between the current amino acid and the average sequence; According to the mutation probability P i,j For each binding region r i Perform adaptive mutation operations, including: For the binding area r i For each amino acid position j in , generate a random number u∈[0,1]; If u≤P i,j , then perform mutation operation on amino acid position j and change r i [j] mutated to a different amino acid (with v i The amino acid distribution in is used as a reference); If u>P i,j , then retain the original amino acid r i [j] unchanged; The mutated binding region is denoted as r i ′, forming a new binding region sequence set {r i '}; Step 3, constructing a binding feature matrix F of the key binding sites of the target antigen based on the antibody library S; Step 4: Perform specific screening on the antibody library S according to the binding feature matrix F to generate a high-affinity antibody subset S′, wherein the high-affinity antibody subset S′ contains antibody sequences that meet the requirements of the feature matrix F; Step 5: Construct a distributed closed-loop optimization system based on the high-affinity antibody subset S′ to generate the final antibody library S * .

2. The method for generating an antibody library according to claim 1, characterized in that: According to the initial antibody library S init Generate the diversity weight matrix W, including: Identification of initial antibody library S init Each sequence s i , the sequence s i Key amino acid fragments in the binding site region i The set R of all antibody binding sites, R = {r1, r2, ..., r n }; Construct a diversity feature space and define the variability of each binding region as a vector v i , and the variability of each binding region is the vector v i represents the probability of mutation of different amino acids at the binding site, as follows: Where count(X) represents the number of amino acids X in the binding site region r. i The frequency of occurrence, L is the total length of the binding region; Combined with the diversity feature vector v i And the initial antibody library S init The statistical characteristics of different binding regions in the matrix are used to generate the diversity weight matrix W.

3. The method for generating an antibody library according to claim 2, characterized in that: The diversity weight matrix W is normalized before output so that it satisfies ∑ j w i,j =1.

4. The method for generating an antibody library according to claim 1, characterized in that: The binding region {r i '} Randomly select two binding region sequences r a ' and r b The recombination process for sequence crossover recombination follows the following rules: The recombination probability of each binding region is dynamically regulated according to its mutation frequency and structural feature distribution; Ensure that the sequence retains the characteristics of the key binding sites during crossover recombination to avoid loss of affinity potential due to mutation; the diversity regularization term Ω(S) is expressed as follows: Among them, sim(s i ,s j ) is the sequence s i and j , δ is the equilibrium parameter that controls the similarity regularization term. Through the regularization term Ω(S), the mutation and recombination frequencies are dynamically adjusted in the antibody generation model to ensure sufficient diversity between sequences to increase the adaptability of the antibody library to a wide range of antigen binding sites.

5. The method for generating an antibody library according to claim 1, characterized in that: The step 3 specifically includes: Obtain the three-dimensional structural data of the target antigen, extract the surface information of the antigen, and use the surface information of the antigen to form a preliminary feature representation of the target antigen. j ; Wherein, the surface information of the antigen is the surface site representation of the surface area of ​​the target antigen, and the local charge density, polarity value and hydrophobicity coefficient are extracted from the surface site representation to form a preliminary feature representation q j ,q j =(φ j ,ψ j ,ρ j ), φ j Indicates site t j The local charge density of j Indicates site t j The polarity value of j Indicates site t j The hydrophobicity coefficient; The preliminary feature representation q j The target antigen surface feature vector set {q j }, and the target antigen surface feature vector set {q j } Perform cluster analysis and calculate the center point b of the target antigen i and the corresponding eigenvector q i is a representative of the combined area; The feature vector q i Integrate into the combined feature matrix F, expressed as follows: Where φ1 represents the local charge density at site t1, φ2 represents the local charge density at site t2, and φ m Indicates site t m The local charge density of site t1 is ψ1, ψ2 is the polarity value of site t2, and ψ m Indicates site t m The polarity value of ρ1 represents the hydrophobicity coefficient of site t1, ρ2 represents the hydrophobicity coefficient of site t2, and ρ m Indicates site t m The hydrophobicity coefficient; Introducing the regularization term Ω(F) into the binding feature matrix F improves the matching accuracy of the binding feature matrix F, thereby ensuring that the antibody library S can adapt to the changes in the specific binding region during the screening process; By combining the feature matrix F and the feature distribution of the antibody library S, the fitness of the antibody and the target antigen binding site is calculated, which is expressed as follows: Among them, score(s k ,b i ) indicates antibody s k Binding site b i The fitness score, f i,j and k,j Represents binding site b i and antibodies k The match score on a specific feature.

6. The method for generating an antibody library according to claim 5, characterized in that: The regularization term Ω(F) is expressed as follows: Among them, dist(q i ,q j ) is the binding site characteristic q i and q j is the Euclidean distance between the feature vectors, τ is the adjustment coefficient, and m represents the total number of sites.

7. The method for generating an antibody library according to claim 1, characterized in that: The specific screening of the antibody library S according to the key binding site feature matrix F to generate an antibody subset S′ with high affinity specifically includes: According to the structural characteristics of the combined feature matrix F, each antibody sequence s in the antibody library S k Calculate its relationship with each binding site b i The specific match degree, that is, the match score of each antibody sequence k,i ; Among them, the matching score of each antibody sequence is match k,i The lowest, indicating that the antibody sequence s k Binding site b i The characteristics are closest to those of the antigen, meeting the binding requirements of the antigen; According to each antibody sequence s k Match score match k,i and the overall affinity of all binding sites to construct each antibody sequence s k The screening score score k , and introduce the regularization term Ω(s k ,F) equilibrium antibody affinity distribution at different binding sites; According to the filter score score k The antibody library S is sorted, and the top N antibody sequences with the lowest scores are selected to form a high-affinity antibody subset S′.

8. The method for generating an antibody library according to claim 7, characterized in that: The matching score match k,i It is expressed as follows: Among them, f i,j Represents the binding site b in the binding feature matrix F i The jth characteristic value of i , polarity ψ i or hydrophobicity ρ i ;s k,j Represents antibodies k The jth characteristic value in the combined area; ω j is the feature importance weight, which controls the contribution of each feature to the matching score; The screening score k , calculated as follows: Among them, α is the regularization term weight, which balances the influence of the regularization term and the matching score; The regularization term Ω(s k ,F), expressed as follows: Among them, dist(q i ,q j ) is the binding site characteristic q i and q j is the Euclidean distance between the feature vectors, and τ is the adjustment coefficient.

9. The method for generating an antibody library according to claim 1, characterized in that: The distributed closed-loop optimization system is constructed based on the high-affinity antibody subset S′ to generate the final antibody library S * , specifically including: Assign the high-affinity antibody subset S′ to multiple nodes, that is, the antibody subset of each node p is S′ p , where p = 1, 2, …, P, where P is the total number of nodes; Perform a specificity score on each node p k and the diversity regularization term Ω(s k ,F), ensure that the antibody subset reaches the local optimum within the node; After completing local optimization at each node p, the specificity of each antibody is scored k and its binding site characteristicsq k The data are transmitted to the master node, which evaluates the binding diversity of the entire antibody library based on the feedback information and calculates the global diversity score diversity(S′). The master node updates the key parameters in the binding feature matrix F and the diversity weight matrix W based on the scores fed back by each node. After receiving the updated key parameters, each computing node recalculates the specificity score of each antibody in each antibody sequence. k , and for the antibody subset S′ p Rescreen the antibodies in the test set and retain the top N antibodies with the lowest scores; After multiple rounds of closed-loop iterations, the antibody subsets of each node will tend to the global optimum, that is, a more balanced state in terms of specificity and diversity. Finally, the antibody subsets of each node will be merged to form the final antibody library S * .

10. The method for generating an antibody library according to claim 9, characterized in that: The specificity score k It is expressed as follows: Among them, match k,i Antibodies k Binding site b i The specific matching score describes the adaptability of the antibody at different binding sites; Ω(s k ,F) is the regularization term, which mainly controls the characteristic smoothness of the binding region to avoid excessive concentration on a few binding sites; α is the regularization coefficient, which is used to regulate the weight of the regularization term in the overall score; The diversity score (S′) is expressed as follows: Among them, dist(q p,i ,q p,j ) represents the antibody s on node p k and l The Euclidean distance on the binding feature, τ is a balance parameter, which controls the effect of the feature distance on the diversity score; the diversity score diversity(S′) is used to measure the coverage of the antibody subset S′ on the binding region.

Images