Antibody de novo design system and method combining target knowledge and local dynamic structure

Abstract

The application discloses an antibody de novo design system and method combining target knowledge and local dynamic structure, and the system comprises a protein reversible generation module, a hidden space decoupling module, a knowledge-driven sequence evaluation module and a complementarity determining region dynamic docking module; the main working scene of the antibody de novo design scheme is target information-driven antibody generation. After the pathogenic protein is determined and the pathogenic mechanism is clarified, an antibody needs to be designed to combine with the target protein to block its function or promote its degradation, and the application can be used for fast and high-credibility antibody generation. The design system generates biologically active proteins through a flow generation model, uses a global knowledge and local target structure double-driven scheme to optimize specific functional proteins as targets, uses a decoupling module to realize dynamic disturbance of the complementarity determining region to further accelerate the docking process, and comprehensively optimizes the antibody de novo design process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and technology, in particular to an antibody de novo design system and method combining target knowledge and local dynamic structure. BACKGROUND

[0002] Protein molecules transmit life signals in vivo through processes such as binding and catalysis. According to the different molecular structures, they specifically undertake various life activities. Antibodies, as a kind of immunoglobulin, are composed of variable and constant regions, and the structure of the variable region determines the specificity of binding. After the human immune system detects an antigen (such as a foreign pathogen), it can produce an antibody that can specifically bind to the antigen, and through cell lysis or guiding phagocytic cells to phagocytize specific antigens, precise treatment is achieved. How to use antigen target information to design and screen antibody molecules with specific functions, and mass-produce using monoclonal antibody technology, is an important breakthrough in the current fight against diseases such as cancer.

[0003] The traditional monoclonal antibody design method is to obtain antibodies by immunizing and separating in mammals and then screening. This method has large resource consumption and long production cycle. To obtain antibodies with certain affinity, multiple rounds of expression enrichment in mammals or in vitro are required, and affinity maturation technology is used for assistance. In addition, antibodies obtained using non-human mammals have the risk of triggering immune rejection.

[0004] The computational antibody design method is usually based on antibody-antigen docking. This method first performs structure prediction on the candidate sequence, and then calculates the affinity index between the antibody and the antigen through flexible docking software (such as Rosetta SungDock), and screens stable structures with high affinity for in vitro experiments. This type of method uses virtual screening to avoid large-scale immunization experiments, but frequent calls to flexible docking software still result in large time consumption.

[0005] In the prior art, part of the technology uses a language model to model the amino acid sequence, and uses autoregressive or non-autoregressive methods to generate the amino acid sequence. This type of method establishes a mapping relationship between the sequence and the protein function through a functional prediction module, uses terms in the Gene Ontology to measure the protein function, and optimizes the amino acid sequence based on this. This type of method is limited by the functional prediction module, which cannot accurately establish a connection with the real function of the protein, and does not introduce its own structure information with the target, so the functional accuracy of the designed protein sequence needs to be improved.

[0006] Another structure-based de novo protein generation method is based on short peptide template library generation. This method docks known short peptides from a short peptide template library onto target proteins with large throughput, and achieves specific design of short peptides by extracting the public part of the docked templates. This method uses structural information for guidance, but is limited by the use of templates, which limits the number of amino acids in short peptides to no more than 50, limiting the application of this method in the field of antibody generation.

[0007] Traditional antibody design methods have the problems of large consumption of computing resources, low reliability of designed antibodies, need for multiple wet experiments to assist in design, and the like, and there is no full-length antibody design scheme for target points. At present, de novo design methods are limited to short peptide binder design or nanobody design.

[0008] Therefore, a de novo antibody design method with high reliability, fast speed and support for long sequence generation is urgently needed. SUMMARY

[0009] The purpose of the present application is to provide a de novo antibody design system and method combining target point knowledge and local dynamic structure, which efficiently samples the conformation of the complementarity determining region on the basis of virtual docking, uses rigid docking software to achieve the effect of flexible docking, and greatly accelerates antibody design. Using global target point knowledge and local docking conditions for double driving, the problem of insufficient specific functionality of designed proteins is improved.

[0010] To achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0011] The present application provides a de novo antibody design system combining target point knowledge and local dynamic structure, comprising: a protein reversible generation module, a hidden space decoupling module, a knowledge-driven sequence evaluation module and a complementarity determining region dynamic docking module;

[0012] The protein reversible generation module is constructed using a flow-based deep generation model, samples from the hidden space encoding of a standard Gaussian distribution, and obtains the sequence and backbone structure of the protein through a reversible neural network;

[0013] The hidden space decoupling module is used to modify the protein structure generated by the flow-based deep generation model, uses a flow parameter to encode the disturbance degree of the complementarity determining region, and after mapping through a multilayer perception network, introduces it into the main network of the flow model through an affine injection layer;

[0014] The knowledge-driven sequence evaluation module uses an existing knowledge graph encoding method to evaluate the function of the protein sequence output by the protein reversible generation module;

[0015] The complementarity determining region dynamic docking module is configured to input the protein sequence and the backbone structure generated by the protein reversible generation module, dock with a target point through a rigid docking software, and output an affinity index of the antibody.

[0016] Further, the antibody de novo design system further comprises:

[0017] The protein molecule dynamics simulation module is configured to perform calculation on the antibody protein by using a molecular dynamics simulation software, control the conformation of the amino acids outside the complementarity determining region to be unchanged, allow the amino acid residues inside the complementarity determining region to be freely disturbed under the action of a molecular force field, simulate the protein fluidity, and use the generated data and the fluidity parameters as training data of the protein reversible generation module and labels of the latent space decoupling module.

[0018] Further, the latent space encoding in the protein reversible generation module comprises two parts, one part is structure decoupling encoding related to the dynamic information of the complementarity determining region, and the other part is a non-decoupling latent variable space related to the overall sequence and structure information of the protein.

[0019] Further, the protein reversible generation module is further configured to receive the data and the fluidity parameters generated by the protein molecule dynamics simulation module as training data, and use the latent space decoupling module to add the decoupling latent variables as conditions to the flow-based deep generation model.

[0020] Further, the latent space decoupling module specifically adopts a decoupling setting similar to StyleGAN, uses part of the variables in the latent space as the complementarity determining region fluidity parameters, and performs constraint on the angle information of the amino acid residues in the complementarity determining region given by the protein molecule dynamics simulation module during training.

[0021] Further, the knowledge-driven sequence evaluation module has the protein sequence generated by the protein reversible generation module as input, and regards the protein function prediction as a multi-label classification problem, uses a protein function predictor pre-trained based on a knowledge graph, captures the relationship between the sequence and the function in the whole function space of the protein by large-scale pre-training based on the knowledge graph, and realizes the function evaluation on the protein sequence.

[0022] In a second aspect, the application further provides an antibody de novo design method combining target point knowledge and local dynamic structure, which uses the antibody de novo design system combining target point knowledge and local dynamic structure according to any one of the above embodiments to realize the training of the flow-based deep generation model and the design of the protein antibody.

[0023] Compared with the prior art, the application has the following beneficial effects:

[0024] The antibody de novo design system provided by the embodiment of the present application combines target knowledge and local dynamic structure, can quickly generate antibody protein sequences and structures with reasonable spatial structure by sampling the coding space, introduces a knowledge graph to realize global screening of sequence functions, realizes detailed search of antibody complementarity determining regions through a reversible flow model, and further screens antibodies through affinity based on docking results. Compared with traditional antibody design methods, the system has the characteristics of high speed, high accuracy, high structural rationality, etc. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The antibody de novo design system provided by the embodiment of the present application combines target knowledge and local dynamic structure, can quickly generate antibody protein sequences and structures with reasonable spatial structure by sampling the coding space, introduces a knowledge graph to realize global screening of sequence functions, realizes detailed search of antibody complementarity determining regions through a reversible flow model, and further screens antibodies through affinity based on docking results. Compared with traditional antibody design methods, the system has the characteristics of high speed, high accuracy, high structural rationality, etc.

[0026] Figure 2 The structure diagram of the protein reversible generation module provided by the embodiment of the present application is shown in the figure.

[0027] Figure 3 The structure diagram of the hidden space decoupling module provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0028] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application is further described below in combination with specific embodiments.

[0029] In the description of the present application, it should be noted that the terms "upper", "lower", "inner", "outer", "front end", "rear end", "two ends", "one end", "the other end" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0030] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "provided with", "connection" and the like should be understood broadly, for example, "connection" can be fixed connection, can also be detachable connection, or integrally connected; can be mechanical connection, can also be electrical connection; can be directly connected, can also be indirectly connected through an intermediate medium, can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0031] According to the protein folding theory, the structure of a protein is uniquely determined by the amino acid sequence, which is in the lowest energy stable conformation in the natural state. In the task of de novo protein design, it is necessary to screen in the functional sparse amino acid sequence combination space and stably bind to a specific target. By regarding the amino acid sequence as a sentence and the protein structure as three-dimensional point cloud data, training a flow-based deep generative model can realize fast generation of proteins through encoding sampling, and can realize de novo generation of sequences and structures without relying on any protein structure starting point or template.

[0032] Referring to Figure 1 The antibody de novo design system provided by the application combines target knowledge and local dynamic structure, and comprises a protein reversible generation module, a protein molecular dynamics simulation module, a latent space decoupling module, a knowledge-driven sequence evaluation module, and a complementarity determining region dynamic docking module.

[0033] The protein reversible generation module is constructed by a flow-based deep generative model, which samples from the latent space encoding of a standard Gaussian distribution and obtains the sequence and backbone structure of the protein through a reversible neural network.

[0034] The protein molecular dynamics simulation module is used to perform operations on antibody proteins using existing molecular dynamics simulation, control the conformation of the amino acid outside the complementarity determining region to be unchanged, observe the changes of the complementarity determining region within a certain time range, and use the generated data and the flow parameters as the training data of the protein reversible generation module and the label of the latent space decoupling module.

[0035] The latent space decoupling module comprises complementarity determining region flowability parameters, a decoupling mapping layer and an affine injection layer. The module modifies the protein structure generated by the flow-based deep generative model, encodes the disturbance degree of the complementarity determining region using the flowability parameters, realizes decoupling through a multi-layer perception network mapping, and then introduces it into the main network of the flow model through the affine injection layer.

[0036] The knowledge-driven sequence evaluation module uses an existing knowledge graph encoding method to evaluate the function of the protein sequence output by the protein reversible generation module.

[0037] The complementarity determining region dynamic docking module accepts the protein sequence and backbone structure generated by the protein reversible generation module, performs docking with the target through a rigid docking software (such as Equidock), and outputs the affinity index of the antibody. This module docks the antibody with multiple complementarity determining region structures obtained by the operation of the latent space decoupling module with the target protein, and comprehensively evaluates the docking affinity in the protein flow background according to the docking scores.

[0038] The antibody design scheme proposed in the embodiment of the application is mainly used in target information driven antibody generation. After the pathogenic protein is determined and the pathogenic mechanism is clarified, an antibody needs to be designed to bind to the target protein to block its function or promote its degradation. The design system proposed in the application can be used to generate antibodies quickly and with high reliability. The design system generates biologically active proteins through a flow generation model, uses a global knowledge and local target structure dual driving scheme to optimize the target protein with specific functions, uses a hidden space decoupling module to further accelerate the docking process by dynamically disturbing the complementarity determining region, and comprehensively optimizes the de novo design process of the antibody.

[0039] The antibody de novo design system combining target knowledge and dynamic structure proposed in the embodiment of the application is a de novo design scheme for pathogenic proteins or signal transduction pathways, which comprehensively uses global knowledge information and local structure docking evaluation to quickly design long sequences with high overall reliability. The system mainly comprises a protein reversible generation module, a protein molecular dynamics simulation module, a hidden space decoupling module, a knowledge driven sequence evaluation module and a complementarity determining region dynamic docking module. The protein reversible generation module uses a flow-based reversible generation model to realize efficient bidirectional information flow between the decoding space and the protein sequence and structure, which is the core innovation point of the application. The specific module composition and function are introduced as follows.

[0040] ① Protein reversible generation module: the module is constructed by using a flow-based deep generation model, and a schematic diagram of the module is shown in FIG. 1. Figure 2 The reversible neural network is constructed by using a flow generation model, and the training data is the sequence and three-dimensional backbone structure information of the protein. Through the forward encoding process (from right to left) of the reversible neural network, the protein data can be mapped to the hidden space encoding. The hidden space encoding comprises two parts, one part is the structure decoupling coding related to the dynamic information of the complementarity determining region, and the other part is the non-decoupling hidden variable space related to the overall sequence and structure information of the protein. The module receives the data generated by the protein molecular dynamics simulation module as the training data, uses the hidden space decoupling module to add the decoupling hidden variables as the condition to the flow-based deep generation model, and in the protein generation process, the sequence and structure information of the protein can be obtained by sampling from the hidden space and through the reverse generation process (from left to right).

[0041] The flow-based deep generation model is used to design a protein backbone with high designability, avoiding the large resource consumption caused by frequent use of protein structure prediction software and docking software. Through training of the antibody structure and sequence in the data set, the deep generation model can learn the distribution of the protein sequence and structure and the relationship therebetween, and only needs to sample from the hidden coding space to quickly obtain the newly generated protein sequence and structure.

[0042] ② The latent space decoupling module includes a complementarity determination region flow parameter, a decoupling mapping layer, and an affine injection layer. A schematic diagram of the module is attached. Figure 3 As shown, this module employs a decoupling setup similar to StyleGAN, using some variables in the latent space as fluidity parameters of the complementarity-determining region (CDR). During training, these parameters are not constrained by a Gaussian distribution, but rather by the amino acid residue angle information of the CDR provided by the protein molecular dynamics simulation module. This parameter set is decoupled through a decoupling mapping layer composed of a multilayer perceptron and then injected additively into the affine coupling layer parameters of the flow generation model via an affine injection layer. For the forward encoding process, leveraging the multi-scale characteristics of the flow generation model, information separated at a larger scale is used as fluidity parameters. This allows the previously separated fluidity parameters to be decoupled and used as conditions for affine injection in subsequent encoding processes.

[0043] ③ The protein molecular dynamics simulation module uses molecular dynamics simulation software (such as Gromacs) to perform calculations on the antibody protein. During the calculation, the conformation of amino acids outside the complementarity-determining region (CDM) is kept unchanged, while the amino acid residues inside the CDM are allowed to move freely under the influence of the molecular force field to simulate the fluidity of proteins under real-world conditions. The simulation lasts for a certain duration, and the protein conformation during the simulation is recorded at a certain step size. The average conformation is taken as the reference origin for the structural fluidity parameters. The changes in the dihedral angles of amino acids in other conformations relative to the average conformation are extracted, normalized, and used as fluidity parameters; these parameters also serve as tags for the implicit space decoupling module.

[0044] ④ The knowledge-driven sequence evaluation module uses the protein sequences generated by the reversible protein generation module as input. It treats protein function prediction as a multi-label classification problem, mapping protein sequences to terms defined by Gene Ontology (GO). The function prediction process uses a pre-trained protein function predictor based on a knowledge graph (such as OntoProtein). This predictor, through large-scale training with the knowledge graph, captures the relationship between sequence and function in the full functional space of proteins. Note that this module only maps protein sequences to existing terms in GO, resulting in a significant deviation from the actual protein function; therefore, it is only used for preliminary screening.

[0045] 5. The input of the dynamic complementarity determining region docking module is the protein backbone structure and sequence generated by the protein reversible generation module. First, the all-atom structure corresponding to the sequence is predicted by a structure prediction software (such as AlphaFold2), the flow parameter is adjusted, the protein backbone structure group with dynamic changes of the complementarity determining region structure under the same sequence is obtained, the relative change amount of the dihedral angle is calculated, and the corresponding adjustment is made on the all-atom structure. The protein all-atom structure group generated in this way is subjected to energy index calculation by a rigid docking software, representing the tightness and affinity of the binding. The protein group with dynamic changes of the complementarity determining region obtained by adjusting the flow parameter is subjected to rigid docking with the target, realizing the sampling process of flexible docking, improving the prediction accuracy, and greatly saving the time cost.

[0046] The global knowledge and target structure co-driving design means proposed in the embodiment of the application effectively solves the problem of low reliability of designed antibodies. On the basis of predicting protein function using a language model, the application further introduces a docking discrimination method, uses a step-by-step screening method to narrow the antibody candidate space, and improves functionality and specificity.

[0047] The flow parameter setting and complementarity determining region dynamic docking scheme proposed in the embodiment of the application further saves time cost. For macromolecular docking tasks, the flexible docking scheme has high precision but large time cost. The application realizes the dynamic changes of the antibody complementarity determining region by modifying the flow parameter, and takes the average affinity index by a rigid docking software, which takes into account the precision and time cost.

[0048] In summary, the application makes targeted improvements in time resources, computing power resources, and wet experimental labor resources, and realizes and accelerates the process of de novo antibody design.

[0049] Based on the same inventive concept, the application also provides a de novo antibody design method combining target knowledge and local dynamic structure, which uses the de novo antibody design system combining target knowledge and local dynamic structure as described in the above embodiment to realize the training of a flow-based deep generation model and the design of a protein antibody.

[0050] Among them:

[0051] 1. The main steps of the model training process are:

[0052] Dynamic data generation-flowing batch training-add mask to strengthen data correlation training three parts constitute, involve module mainly protein reversible generation module, protein molecular dynamics simulation module and hidden space decoupling module. The data set contains protein structure with a determined conformation and the corresponding sequence. First, the range of the complementarity determining region (CDR region) is obtained according to the antibody coding rule, one light chain and one heavy chain of the antibody are connected to form a single-chain antibody (scFv), the single-chain antibody is processed by the protein molecular dynamics simulation module, the conformation outside the CDR region is controlled to be unchanged, and the conformation inside the CDR region is disturbed, and the generated protein structure data is stored in batches, and the flowing parameter tag is obtained according to the protein molecular dynamics simulation module.

[0053] The training process is that after the main structure of the protein molecule is extracted, the main structure is input into the flow generation model together with the sequence, and the hidden space code is obtained through the forward encoding module. The undecoupled part loss function is the flow generation model loss function, and the decoupled part loss function is the mean square error calculated with the flowing parameter tag. In the mask strengthening training process, the protein main structure is subjected to mask processing when inputting into the flow model, so as to emphasize the self-consistency of the sequence and structure generated by the generation model.

[0054] 2. The main steps of the protein design process are:

[0055] Hidden space sampling-protein sequence structure generation-sequence functional evaluation-structure affinity evaluation, involve module mainly protein reversible generation module, hidden space decoupling module, knowledge driven sequence evaluation module and complementarity determining region dynamic docking module. First, the protein sequence and structure are obtained through hidden space sampling, at this time the flowing parameter is 0, representing the average conformation. The generated sequence is functionally evaluated and subjected to the first round of screening. The sequence screened is subjected to other structure prediction software to obtain its protein all-atom structure, and the similarity with the main structure obtained by the generation model is used for the second round of screening. The flowing parameter is modified to obtain the change amount of the protein main structure, and the all-atom structure is modified synchronously to obtain a protein structure group, the structure group is subjected to rigid docking with the target and the average docking energy is taken to perform the third round of screening.

[0056] The antibody de novo design method of the application combined with target knowledge and local dynamic structure is described in detail below: the antibody design process targeting human epidermal growth factor receptor-2 (HER2) protein:

[0057] The HER2 gene is an important target of human breast cancer tumor targeted therapy drugs. Overexpression of HER2 in tumors promotes the proliferation of cancer cells and inhibits apoptosis, and the design of antibodies targeting the binding of HER2 protein can effectively inhibit the proliferation of cancer cells.

[0058] The antibody structure database (SAbDab: The Structural Antibody Database) is selected as the data set, the data is expanded using the molecular dynamics simulation software GROMACS, the protein conformation sequence obtained by simulation is averaged, and the flow parameter is calculated according to the average conformation as the label information of the hidden space decoupling module, and the protein reversible generation module is trained.

[0059] After training convergence, Monte Carlo sampling is performed in the hidden space to obtain generated protein sequences and structures. The generated samples are screened in the first round by the knowledge-driven sequence evaluation module, and the GO terms related to HER2 targeting antibodies such as "immunoglobulin", "human epidermal growth factor receptor-2 binder" and the like are selected, and a certain threshold screening is performed. In the screened antibody samples, the 3D structure of the whole atom of the antibody is predicted by using the sequence, and compared with the structure obtained by the generation model, the reliability is evaluated and a further round of screening is performed. The candidate antibodies screened are adjusted to form an antibody group by adjusting the flow parameter, and are rigidly docked with the HER2 protein in batches to obtain the average affinity index, and the antibody sequences with high ranking are screened as the output of the algorithm design, which is used for subsequent wet experiment verification.

[0060] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. An antibody de novo design system combining target knowledge and local dynamic structure, characterized in that, The method comprises the following steps: The protein reversible generation module, the hidden space decoupling module, the knowledge-driven sequence evaluation module, and the complementarity determining region dynamic docking module; The protein reversible generation module is constructed by using a flow-based deep generation model, samples from a hidden space coding of a standard Gaussian distribution, and obtains the sequence and backbone structure of the protein through a reversible neural network; the hidden space coding comprises two parts; one part is structure decoupling coding related to dynamic information of the complementarity determining region; the other part is a non-decoupling hidden variable space related to overall sequence and structure information of the protein; The hidden space decoupling module is used for modifying the protein structure generated by the flow-based deep generation model, using a flow parameter to encode the disturbance degree of the complementarity determining region, and introducing the complementarity determining region into the main network of the flow model through a multi-layer perception network mapping and an affine injection layer after decoupling; The knowledge-driven sequence evaluation module uses an existing knowledge graph coding method to evaluate the function of the protein sequence output by the protein reversible generation module; The complementarity determining region dynamic docking module is used for inputting the protein sequence and backbone structure generated by the protein reversible generation module, performing docking with a target through a rigid docking software, and outputting an affinity index of the antibody; The protein molecular dynamics simulation module uses a molecular dynamics simulation software to perform calculation on the antibody protein, controls the conformation of the amino acids outside the complementarity determining region to be unchanged, allows the amino acid residues inside the complementarity determining region to freely disturb under the action of a molecular force field, simulates the flowability of the protein, and uses the generated data and the flow parameter as training data of the protein reversible generation module and a label of the hidden space decoupling module; The hidden space decoupling module specifically uses a decoupling setting similar to StyleGAN, uses part of the variables of the hidden space as the flowability parameter of the complementarity determining region, and is constrained by the angle information of the amino acid residues of the complementarity determining region given by the protein molecular dynamics simulation module during training; The knowledge-driven sequence evaluation module uses the protein sequence generated by the protein reversible generation module as input, regards protein function prediction as a multi-label classification problem, uses a protein function predictor pre-trained based on a knowledge graph, captures the relationship between the sequence and the function in the whole function space of the protein through large-scale pre-training based on the knowledge graph, and realizes function evaluation on the protein sequence.

2. The antibody de novo design system of claim 1, wherein, The protein reversible generation module is also used for receiving the data and the flow parameter generated by the protein molecular dynamics simulation module as training data, and adding the decoupled hidden variables as conditions to the flow-based deep generation model by using the hidden space decoupling module.

3. Antibody de novo design method combining knowledge of target sites and local dynamic structure, characterized in that, The antibody de novo design system combining target knowledge and local dynamic structure is used for training the flow-based deep generation model and designing the protein antibody.

Images