System for designing and optimizing D-type biomolecules and method and application thereof
The generation of D-type biomolecule binding skeletons through stereoisomeric conversion and evolution Bayesian optimization algorithms solves the problem of low design efficiency in traditional methods, realizes efficient combination design and optimization, and improves the binding ability of D-type biomolecule.
Patent Information
- Application Number
- CN202510194890.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-21
- Publication Date
- 2025-07-18
AI Technical Summary
The prior art is difficult to quickly and efficiently design D-type biomolecules that can bind to the target protein with high specificity and high affinity, and traditional methods rely on time-consuming molecular simulation and docking technology to meet design needs.
The stereoisomeric conversion module, the binding skeleton generation module, the sequence design module and the evolution Bayesian optimization algorithm are used to generate the binding skeleton through stereoisomeric conversion, and the adaptive amino acid sequence is generated using deep learning and probability models. The sequence is iteratively optimized through the evolution Bayesian optimization algorithm to improve binding affinity and specificity.
It significantly improves the binding affinity of D-type biomolecules with target proteins, has high optimization efficiency, is suitable for different types of D-type biomolecules, and provides a powerful tool for drug development and biotechnology research.
Smart Images
Figure CN120340622A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of biomolecular design and optimization, and relates to a system, a method and an application for designing and optimizing D-type biomolecules. Specifically, it relates to a method and a system for designing and optimizing biomolecules composed of D-amino acids (including D-peptides, D-oligopeptides and D-proteins), enabling them to efficiently bind to target proteins, and provides its specific applications. Background Art
[0002] Biomolecules composed of D-amino acids have great application potential in the fields of drug development and biotechnology. Compared with natural L-amino acids, D-type biomolecules have higher resistance to protease degradation, and thus have a longer half-life and higher bioavailability in vivo. However, due to the differences in stereochemical properties between D-amino acids and L-amino acids, designing D-type biomolecules that can effectively bind to target proteins has always been a challenge. Traditional methods usually rely on time-consuming molecular simulation and docking techniques, which are difficult to meet the requirements of rapid and efficient design. Summary of the Invention
[0003] The present invention is directed to the above problems, and aims to provide a system (Mirror-Peptidizer system) and a method for designing and optimizing D-type biomolecules, enabling biomolecules composed of D-amino acids to bind to target proteins with high specificity and high affinity. Then, it is verified by designing and optimizing D-type polypeptide conjugates of PDL1 and MDM2.
[0004] In the first aspect of the present invention, a system for designing and optimizing D-type biomolecules is provided, including a stereoisomer conversion module, a conjugate backbone generation module, a sequence design module and an optimization module.
[0005] Among them, the stereoisomer conversion module is used to perform stereoisomer conversion on the target protein and the generated conjugate before initial design and after generating the sequence to obtain D-type biomolecules;
[0006] The conjugate backbone generation module receives the structural information of the stereoisomer target protein and generates a conjugate backbone under structural constraints;
[0007] The sequence design module predicts the amino acid sequence adapted to the conjugate backbone;
[0008] The module iteratively optimizes the sequence based on the evolutionary Bayesian optimization algorithm to improve the binding affinity and specificity of D-type biomolecules.
[0009] The preferred technical solutions of the above modules are as follows:
[0010] (1)The stereoisomer conversion module realizes stereoisomer conversion by performing spatial inversion along the x, y, or z axis, or through plane symmetry operations. Specifically:
[0011] For any point (x, y, z) in space, its inversion operation can be expressed as:
[0012] 。
[0013] By selecting appropriate inversion operations, the stereoisomer structure of the target protein can be obtained for subsequent conjugate design.
[0014] (2)The conjugate backbone generation module
[0015] Constructed based on deep learning or probability models, spatial constraints and energy functions are applied during the generation process to generate the conjugate backbone based on the surface features and spatial conformation of the target protein. The generation model needs to fully consider the surface features and spatial conformation of the target protein to ensure that the generated conjugate can fit the stereoisomer structure of the target protein.
[0016] The generation process includes the following steps:
[0017] Conditional input: The structural information of the stereoisomer target protein is used as conditional input to guide the conjugate generation process;
[0018] Structural constraints: During the generation process, spatial constraints and energy functions are applied to ensure that the generated conjugate backbone meets specific structural requirements and fits the stereoisomer target protein.
[0019] Generation model: Deep learning or probability models (such as diffusion models, variational autoencoders, etc.) are used to generate a conjugate backbone that meets the requirements under structural constraints.
[0020] In the present invention, the diffusion model can be regarded as starting from noise and gradually generating a reasonable binder backbone coordinate by iterative denoising and fitting to the latent distribution ,The conditional probability distribution is expressed as:
[0021]
[0022] Among them, the set of D-type receptor coordinates is ,
[0023] Through this diffusion process, a three-dimensional backbone compatible with the receptor structure is generated for the binder region under the constraint of the D-type receptor:
[0024]
[0025] (3)The sequence design module
[0026] Input the generated conjugate backbone into the sequence design model to predict the compatible amino acid sequence. The sequence design model needs to consider the spatial structure of the backbone and its interaction with the stereoisomeric target protein to generate a sequence with high binding potential.
[0027] The sequence prediction methods include the following two:
[0028] Energy function optimization: Define the energy function E(x) to evaluate the fitness of the sequence to the backbone, and find the optimal sequence by minimizing the energy function.
[0029] Probability model: Utilize the conditional probability model (such as the Bayesian classification model) to calculate the probability distribution of the amino acid sequence given the backbone structure, and select the sequence with the highest probability.
[0030] (4) Optimization module
[0031] Used to further improve the binding affinity and specificity of D-type biomolecules, and the Evolutionary Bayesian Optimization (EBO) algorithm is used to iteratively optimize its sequence. This algorithm constructs a surrogate model to predict the binding affinity and uncertainty of the sequence, and uses the acquisition function to intelligently explore in the sequence space, select potential high-affinity sequences for evaluation and update.
[0032] Specifically, the optimization module includes an initialization sub-module and an iterative optimization sub-module.
[0033] The initialization sub-module uses the generated D-type biomolecule sequence as the initial data set , where is the true binding affinity of the sequence; and based on the initial data set, train the surrogate model f(x) to predict the mean binding affinity and uncertainty of the sequence.
[0034] Preferably, according to the data scale, select the Gaussian Process (GP) or Bayesian Neural Network (BNN) as the surrogate model.
[0035] The process of the iterative optimization sub-module is as follows:
[0036] Candidate sequence generation: On the specified site set L, generate a candidate sequence set through mutation and recombination operations;
[0037] Acquisition function evaluation: For each candidate sequence , calculate the acquisition function value ; Preferably, according to the optimization requirements, select the Upper Confidence Bound (UCB) or Expected Improvement (EI) as the acquisition function.
[0038] Sequence selection and evaluation: Select the sequence with the largest acquisition function value , and perform a true binding affinity evaluation on it to obtain ;
[0039] Surrogate model update: Add the new data point to the dataset D and update the surrogate model ;
[0040] Convergence judgment: If the maximum number of iterations is reached or the preset convergence condition is satisfied, stop the optimization.
[0041] Among them, in the candidate sequence generation step, the site set L can be specified in any of the following ways:
[0042] (1) Allow the user to specify the set of mutable amino acids for each site. For example, in order to avoid the influence of certain amino acids, it is possible to freely set that cysteine or other specific amino acids are not included when mutating the site, so as to reduce the possibility of generating infeasible sequences.
[0043] (2) Use strategies such as Thompson Sampling to select sites with greater potential for improving binding affinity for mutation according to the prediction uncertainty of the current model, so as to optimize the sequence exploration efficiency.
[0044] (3) Use the conditional probability model trained by the sequence prediction module to calculate the conditional probability distribution of each site, and preferentially select the mutation direction according to the amino acid with the highest probability, or only generate candidate sequences for sites with lower probabilities to enhance the pertinence of the optimization.
[0045] The second aspect of the present invention provides a method for designing and optimizing D-type biomolecules, including the following steps:
[0046] A. Initial conjugate design
[0047] (1) Obtain the three-dimensional structure of the target protein: Generate its stereoisomer structure by performing spatial inversion along a certain coordinate axis (x, y or z axis), or by plane symmetry operation to perform stereoisomer transformation on the target protein; By selecting appropriate inversion operations, the stereoisomer structure of the target protein can be obtained for subsequent conjugate design.
[0048] (2) Constrained conjugate generation: After obtaining the stereoisomer target protein structure, apply spatial constraints and energy functions, fully consider the surface characteristics and spatial conformation of the target protein, and use deep learning or probability models (such as diffusion models, variational autoencoders) to generate a constrained conjugate backbone under its structural constraints.
[0049] (3) Sequence prediction: Based on the spatial structure of the scaffold and its interaction with the target protein of the stereoisomer, the sequence design module predicts the amino acid sequence based on an energy function or a probability model. When using an energy function, first define the energy function E(x) to evaluate the fitness of the sequence to the conjugate scaffold, and find the optimal sequence by minimizing the energy function. When using a probability model, use the conditional probability model to calculate the probability distribution of the amino acid sequence given the scaffold structure, and select the sequence with the highest probability;
[0050] (4) Stereoisomer restoration: Perform stereoisomer conversion on the target protein of the stereoisomer and the generated conjugate again to restore to the original target protein structure and the corresponding D-type biomolecular conjugate. The obtained D-type biomolecule can effectively bind to the target protein.
[0051] B. Conjugate design optimization
[0052] Use the initialization sub-module to use the generated D-type biomolecular sequence as the initial data set where is the true binding affinity of the sequence; and train a surrogate model based on the initial data set to predict the mean binding affinity and uncertainty of the sequence;
[0053] Use the iterative optimization sub-module for iterative optimization. The process is as follows:
[0054] Candidate sequence generation: Generate a set of candidate sequences through mutation and recombination operations on the specified site set L ;
[0055] Acquisition function evaluation: For each candidate sequence , calculate the acquisition function value ;
[0056] Sequence selection and evaluation: Select the sequence with the largest acquisition function value , and perform a true binding affinity evaluation on it to obtain ;
[0057] Surrogate model update: Add the new data point to the data set D and update the surrogate model ;
[0058] Convergence judgment: If the maximum number of iterations is reached or the preset convergence condition is satisfied, stop the optimization.
[0059] The third aspect of the present invention provides an application of the above-described system, which relates to a specific application in the design and optimization of D-peptide conjugates of PDL1 and MDM2.
[0060] In the application in the design and optimization of D-peptide conjugates of PDL1, the initial sequence is TPWALVTDGGGVLNL (SEQ ID NO.2), and its binding constant ( ) is 313 nM; after optimization, the amino acid sequence of the D-peptide conjugate is TPWSLVTDGGGVLNL (SEQ ID NO.1), and the binding constant is 64.9 nM, with the binding force increased by about 5 times.
[0061] In the application in the design and optimization of D-peptide conjugates of MDM2, after preliminary calculation and screening, 9 most promising MBDP candidate sequences were finally selected for subsequent experimental verification. The results showed that 4 out of the 9 MBDP candidate molecules exhibited high binding affinity (see Table 2). Among them, MBDP-1 performed the best, and its with MDM2 reached a high level of about 11.9 nM. The K_D values of MBDP-2, MBDP-3, and MBDP-4 were 86 nM, 584 nM, and 876 nM, respectively.
[0062] The innovations of the present invention are as follows:
[0063] Combination of stereoisomer conversion and structural constraint: By performing stereoisomer conversion on the target protein and generating conjugates under its structural constraints, the spatial conformation fitness of the generated D-type biomolecules is ensured, and the effectiveness of the initial design is improved.
[0064] Application of the evolutionary Bayesian optimization algorithm: An advanced machine learning algorithm is introduced, and the surrogate model and acquisition function are used to intelligently explore and optimize in the sequence space, improving the optimization efficiency and accuracy.
[0065] Automation and generality of the system: Each module is integrated into an automated computational process. From target protein preparation, conjugate generation, sequence prediction, stereoisomer restoration to D-type biomolecule optimization, it does not depend on specific third-party tools and has high generality and scalability.
[0066] Through the above method, the present invention can effectively design and optimize D-type biomolecules, enabling them to possess high binding ability and good biological properties, providing a powerful tool for drug development and biotechnology research.
[0067] Through practical applications, the method of the present invention has achieved remarkable results in the design and optimization of D-peptides, D-oligopeptides, and D-proteins. Specifically:
[0068] Enhanced binding affinity: After optimization, the designed D-type biomolecule shows a significantly improved binding affinity with the target protein, demonstrating the effectiveness of the method.
[0069] Optimization efficiency: Compared with traditional methods, the evolutionary Bayesian optimization algorithm can find high-affinity sequences within fewer iterations, significantly improving the optimization efficiency.
[0070] Wide applicability: This method is applicable to different types of D-type biomolecules and has broad application prospects, providing new solutions for the fields of biopharmaceuticals and bioengineering. Brief Description of the Drawings
[0074] Figure 1 Shows the structural block diagram for designing and optimizing the D-type biomolecule system in the present invention;
[0075] Figure 2 Shows the method flow for designing and optimizing the D-type biomolecule in the present invention. Detailed Description of the Invention
[0071] The following examples and experimental examples further illustrate the present invention and should not be construed as limiting the present invention. The examples do not include a detailed description of traditional methods, which are well-known to those of ordinary skill in the art and have been described in many publications.
[0072] Unless otherwise defined, all professional and scientific terms used herein have the same meaning as those familiar to those skilled in the art. In addition, any method similar or equivalent to the described content can be applied to the present invention, and the preferred implementation methods and materials described in the detailed description are for illustrative purposes only.
[0073] According to Figure 1 , the system 100 for designing and optimizing the D-type biomolecule in this embodiment includes a stereoisomer conversion module 1, a conjugate backbone generation module 2, a sequence design module 3, an optimization module 4, and a control module 5.
[0074] The stereoisomer conversion module 1 is used to perform stereoisomer conversion on the target protein and the generated conjugate before the initial design and after the sequence generation to obtain the D-type biomolecule. Specifically, the stereoisomer conversion can be achieved by performing a spatial inversion along the x, y, or z axis, or by a plane symmetry operation, and the stereoisomer structure of the target protein can be obtained for subsequent conjugate design.
[0075] The conjugate backbone generation module 2 is constructed based on deep learning or probability models. During the generation process, spatial constraints and energy functions are applied to generate the conjugate backbone based on the surface characteristics and spatial conformation of the target protein. The generation model needs to fully consider the surface characteristics and spatial conformation of the target protein to ensure that the generated conjugate can fit the stereoisomer structure of the target protein.
[0076] The sequence design module 3 inputs the generated conjugate backbone into the sequence design model to predict the amino acid sequence that fits it. The sequence design model needs to consider the spatial structure of the backbone and its interaction with the stereoisomer target protein to generate a sequence with high binding potential.
[0077] The optimization module 4 is used to further improve the binding affinity and specificity of the D-type biomolecule. The evolutionary Bayesian optimization (EBO) algorithm is used to iteratively optimize its sequence. This algorithm constructs a surrogate model to predict the binding affinity and uncertainty of the sequence, and uses the acquisition function to intelligently explore in the sequence space to select potential high-affinity sequences for evaluation and update.
[0078] Specifically, the optimization module 4 includes an initialization sub-module 41 and an iterative optimization sub-module 42.
[0079] The initialization sub-module 41 uses the generated D-type biomolecule sequence as the initial data set , where is the true binding affinity of the sequence; and a surrogate model is trained based on the initial data set to predict the mean binding affinity and uncertainty of the sequence.
[0080] Preferably, according to the data scale, a Gaussian process (GP) or a Bayesian neural network (BNN) is selected as the surrogate model.
[0081] The process of the iterative optimization sub-module 42 is as follows:
[0082] Candidate sequence generation: On the specified site set L, a candidate sequence set is generated through mutation and recombination operations X cand ;
[0083] Acquisition function evaluation: For each candidate sequence , the acquisition function value is calculated; Preferably, according to the optimization requirements, the upper confidence bound (UCB) or the expected improvement (EI) is selected as the acquisition function.
[0084] Sequence selection and evaluation: Select the sequence with the largest acquisition function value , and perform a true binding affinity evaluation on it to obtain ;
[0085] Surrogate model update: Add new data points to the dataset D and update the surrogate model ;
[0086] Convergence judgment: If the maximum number of iterations is reached or the preset convergence condition is satisfied ( ), stop the optimization.
[0087] According to Figure 2 , a method for designing and optimizing D-type biomolecules includes the following steps:
[0088] A. Initial conjugate design
[0089] (1) Obtain the three-dimensional structure of the target protein: Generate its stereoisomer structure by performing spatial inversion along a certain coordinate axis (x, y, or z axis) or through plane symmetry operations on the target protein; By selecting appropriate inversion operations, the stereoisomer structure of the target protein can be obtained for subsequent conjugate design.
[0090] (2) Constrained conjugate generation: After obtaining the stereoisomer structure of the target protein, apply spatial constraints and energy functions, fully considering the surface characteristics and spatial conformation of the target protein, and use deep learning or probability models (such as diffusion models, variational autoencoders) to generate a constrained conjugate backbone under its structural constraints.
[0091] (3) Sequence prediction: The sequence design module predicts the amino acid sequence based on the spatial structure of the backbone and its interaction with the stereoisomer target protein, either based on an energy function or a probability model; When based on an energy function, first define the energy function E(x) to evaluate the fitness of the sequence to the conjugate backbone, and find the best sequence by minimizing the energy function; When based on a probability model, use the conditional probability model to calculate the probability distribution of the amino acid sequence given the backbone structure and select the sequence with the highest probability;
[0092] (4) Stereoisomer restoration: Perform stereoisomer transformation on the stereoisomer target protein and the generated conjugate again to restore to the original target protein structure and the corresponding D-type biomolecule conjugate. The obtained D-type biomolecule can effectively bind to the target protein.
[0093] B. Conjugate design optimization
[0094] Implemented based on the evolutionary Bayesian optimization algorithm.
[0095] First, use the initialization sub-module to use the generated D-type biomolecular sequence as the initial dataset , where is the true binding affinity of the sequence; and train the surrogate model based on the initial dataset to predict the mean binding affinity and uncertainty of the sequence;
[0096] Then, use the iterative optimization sub-module for iterative optimization, and the process is as follows:
[0097] Candidate sequence generation: On the specified site set L, generate a candidate sequence set through mutation and recombination operations ;
[0098] Acquisition function evaluation: For each candidate sequence , calculate the acquisition function value ;
[0099] Sequence selection and evaluation: Select the sequence with the largest acquisition function value , and perform a true binding affinity evaluation on it to obtain ;
[0100] Surrogate model update: Add the new data point to the dataset D and update the surrogate model ;
[0101] Convergence judgment: If the maximum number of iterations is reached or the preset convergence condition ( ) is satisfied, then stop the optimization.
[0102] Specifically, the steps of the conjugate design optimization algorithm are as follows:
[0103] Input:
[0104] - Target protein structure target_structure;
[0105] - Initial sequence set initial_sequences;
[0106] - Optimization site set L;
[0107] - Amino acid set A;
[0108] - Maximum number of iterations N_max;
[0109] - Convergence threshold .
[0110] Output:
[0111] -Optimal D-Type Biomolecular Sequences 。
[0112] Steps:
[0113]
[0114]
[0115] Application Example 1 Optimizing the D-Type Peptide Conjugate of PDL1 Using the Mirror-Peptidizer Method
[0116] PDL1 is an important immune checkpoint protein, and its regulatory mechanism plays a crucial role in cancer immunotherapy. Through the AI-driven Mirror-Peptidizer method, the present invention successfully designed a high-affinity D-type peptide conjugate, and further improved the binding activity through Bayesian optimization.
[0117] The experimental procedure is as follows:
[0118] 1. Initial Conjugate Design
[0119] In the initial design stage (Tier 1) of the Mirror-Peptidizer method, first, the crystal structure of PDL1 (PDB ID: 7OUN) was obtained, stereoisomerically converted to generate its D conformation; subsequently, under its structural constraints, a generative model algorithm was used to generate the backbone of the initial L-type peptide conjugate and predict the compatible amino acid sequence. Through stereoisomeric restoration, the first candidate D-type peptide conjugate was obtained.
[0120] Initial sequence: TPWALVTDGGGVLNL (SEQ ID NO.2).
[0121] Initial activity: Binding constant with PDL1 : 313 nM.
[0122] 2. Design Optimization
[0123] In the optimization stage (Tier 2), using the evolutionary Bayesian optimization algorithm described above, based on the initial sequence, the sequence space was intelligently explored through mutation and screening. Each iteration was based on the surrogate model to predict the binding affinity, and the acquisition function was used to select the optimal candidate sequence for evaluation.
[0124] The optimization strategy adopted is as follows:
[0125] Optimization goal: Introduce mutations at sequence positions to find amino acid variants that can improve the binding affinity. Selected optimization site: A4. Set of mutable amino acids: The standard 20 amino acids.
[0126] Optimization result: After several rounds of optimization, mutation S (Serine) was selected at site A4. 。 Finally, the optimized D-type polypeptide sequence was obtained:
[0127] Optimized sequence: TPWSLVTDGGGVLNL (SEQ ID NO.1);
[0128] Optimized activity: Binding constant with PDL1 : 64.9 nM.
[0129] The experimental results are shown in Table 1 below, indicating that through the Mirror-Peptidizer method, the designed initial conjugate already has a relatively high binding ability (313 nM). 。 After Bayesian optimization, the sequence was further finely tuned, and the binding ability was successfully increased to 64.9 nM, improving the binding affinity by about 5 times.
[0130] Table 1 Comparison of D-type polypeptide conjugates of PDL1:7OUN before and after optimization
[0131] Application Example 2: Design and verification of a D-type polypeptide conjugate targeting MDM2 using the Mirror-Peptidizer method
[0132] MDM2 is an important negative regulator in the p53 pathway and has a key impact on the proliferation and survival of various tumor cells. By perturbing the interaction between MDM2 and p53, p53 activity can be restored, thereby inducing apoptosis of tumor cells. Therefore, designing D-type polypeptides (MBDPs) that can effectively bind to MDM2 has potential value in anti-tumor drug development.
[0133] Using the Mirror-Peptidizer method of the present invention, with MDM2 as the target protein, its structure was stereoisomerically converted to generate an artificially constructed D-type conformation as a constraint condition. Subsequently, a generative model was used to generate an initial L-type polypeptide backbone under this constraint condition, and an amino acid sequence adapted to this structure was designed through a sequence prediction module. Then, the obtained conjugate was stereoisomerically restored to obtain the final D-type polypeptide candidate (MBDPs).
[0134] After preliminary calculation and screening, a total of 9 most promising MBDP candidate sequences were selected for subsequent experimental verification, as shown in Table 2 below:
[0135] Table 2: Summary of sequences and dissociation constants of D-type polypeptide conjugates MBDP-1 to MBDP-9 of MDM2
[0136]
[0137] To verify the actual binding effect of the MBDPs designed by Mirror-Peptidizer, we performed solid-phase peptide synthesis (SPPS) and purification on the above 9 MBDP candidate molecules. The binding affinity (dissociation constant ) with MDM2 was determined using isothermal titration calorimetry (ITC).
[0138] The ITC results showed that 4 out of the 9 MBDP candidate molecules exhibited high binding affinity. Among them, MBDP-1 performed the best, with a reaching a high level of approximately 11.9 nM with MDM2. The K_D values of MBDP-2, MBDP-3, and MBDP-4 were 86 nM, 584 nM, and 876 nM, respectively. Notably, by comparing the binding activity of MBDPs with the scores of MPNN (Molecular Property Prediction Neural Network) used in the present invention, a high correlation (logarithmic correlation coefficient = 0.9562) was found, indicating good consistency between the computational model constructed by Mirror-Peptidizer and the experimental results.
[0139] The parts not described in the present invention are the same as or implemented using the prior art. The applicant declares that the present invention illustrates the detailed method of the present invention through the above embodiments, but the present invention is not limited to the above detailed method, that is, it does not mean that the present invention must rely on the above detailed method to be implemented. Those skilled in the art should understand that any improvement to the present invention, the equivalent substitution of each raw material of the products of the present invention, the addition of auxiliary components, and the selection of specific methods, etc., all fall within the protection scope and disclosure scope of the present invention.
Claims
1. A system for designing and optimizing D-type biomolecules, characterized in that, It includes a stereoisomer conversion module, a conjugate backbone generation module, a sequence design module, and an optimization module. Among them, the stereoisomer conversion module is used to perform stereoisomer conversion on the target protein and the generated conjugate before the initial design and after the sequence generation to obtain D-type biomolecules. The conjugate backbone generation module receives the structural information of the stereoisomeric target protein and generates a conjugate backbone under structural constraints. The sequence design module predicts the amino acid sequence adapted to the conjugate backbone. The optimization module iteratively optimizes the sequence based on the evolutionary Bayesian optimization algorithm to improve the binding affinity and specificity of the D-type biomolecule.
2. The system for designing and optimizing D-type biomolecules according to claim 1, characterized in that, The stereoisomer conversion module realizes stereoisomer conversion by performing spatial inversion along the x, y, or z axis, or by plane symmetry operation.
3. The system for designing and optimizing D-type biomolecules according to claim 1, wherein The conjugate backbone generation module is constructed based on deep learning or a probability model, and applies spatial constraints and energy functions during the generation process to generate a conjugate backbone based on the surface features and spatial conformation of the target protein.
4. The system for designing and optimizing D-type biomolecules according to claim 3, characterized in that, The probability model is selected from diffusion models, which start from noise and gradually generate reasonable binder backbone coordinates by iteratively denoising to fit the latent distribution. , and the conditional probability distribution is expressed as: Among them, is the set of D-type receptor coordinates, Through this diffusion process, a tri- is generated for the binder region that is compatible with the receptor structure under the given D-type receptor constraint. Main chain backbone : 。 5. The system for designing and optimizing D-type biomolecules according to claim 1, characterized in that, The sequence design module predicts the amino acid sequence based on an energy function or a probability model. When based on the energy function, first define the energy function E(x) to evaluate the fitness of the sequence to the binder backbone, and find the optimal sequence by minimizing the energy function; when based on the probability model, use the conditional probability model , calculate the probability distribution of the amino acid sequence under the given backbone structure, and select the sequence with the highest probability.
6. The system for designing and optimizing D-type biomolecules according to claim 1, characterized in that, The optimization module includes an initialization sub-module and an iterative optimization sub-module. The initialization sub-module uses the generated D-type biomolecular sequence as the initial data set , where is the true binding affinity of the sequence; and training a surrogate model based on the initial dataset , predicting the mean binding affinity of the sequence and uncertainty ; The process of the iterative optimization sub-module is as follows: Candidate sequence generation: On the specified site set L, a candidate sequence set is generated through mutation and recombination operations X cand ; Acquisition function evaluation: For each candidate sequence , calculate the acquisition function value ; Sequence selection and evaluation: Select the sequence x with the largest acquisition function value and evaluate its true binding affinity to obtain next , and obtain ; Surrogate model update: Add the new data point to the dataset D and update the surrogate model ; Convergence judgment: If the maximum number of iterations is reached or the preset convergence condition is satisfied, the optimization is stopped.
7. The system for designing and optimizing D-type biomolecules according to claim 6, wherein In the candidate sequence generation step, the set of sites L is specified by any of the following methods: (1) Allow the user to specify the set of mutable amino acids at each site to reduce the possibility of generating infeasible sequences. (2) Use strategies such as Thompson Sampling, and according to the prediction uncertainty of the current model, select the sites with greater potential for improving binding affinity for mutation to optimize the sequence exploration efficiency. (3) Use the conditional probability model trained by the sequence prediction module to calculate the conditional probability distribution of each site, and preferentially select the mutation direction according to the amino acid with the highest probability, or only generate candidate sequences for the sites with lower probabilities. According to the data scale, select the Gaussian process (GP) or the Bayesian neural network (BNN) as the surrogate model f(x). Select the upper confidence bound (UCB) or the expected improvement (EI) as the acquisition function. The convergence condition is set as .
8. A method for designing and optimizing D-type biomolecules, characterized in that, It includes the following steps: A. Initial conjugate design (1) Obtain the three-dimensional structure of the target protein: Perform stereoisomer conversion on the target protein by performing spatial inversion along the x, y, or z axis, or by plane symmetry operation to generate its stereoisomeric structure. (2) Constrained conjugate generation: After obtaining the stereoisomeric target protein structure, apply spatial constraints and energy functions, fully consider the surface features and spatial conformation of the target protein, and generate a constrained conjugate backbone under its structural constraints. (3) Sequence prediction: The sequence design module predicts the amino acid sequence based on an energy function or a probability model; when based on the energy function, first define the energy function , evaluate the fitness of the sequence with the binder backbone, and find the optimal sequence by minimizing the energy function; when based on the probability model, use the conditional probability model , calculate the probability distribution of the amino acid sequence under a given backbone structure, and select the sequence with the highest probability; (4)Stereoisomer restoration: The stereoisomeric target protein and the generated conjugate are subjected to stereoisomeric conversion again to restore to the original target protein structure and the corresponding D-type biomolecule conjugate. The obtained D-type biomolecule can effectively bind to the target protein. B. Conjugate design optimization An initialization sub-module is adopted to use the generated D-type biomolecular sequence as the initial data set , where is the true binding affinity of the sequence; And train a surrogate model based on the initial dataset , to predict the mean binding affinity of the sequence and uncertainty ; Iterative optimization is carried out using the iterative optimization sub-module, and the process is as follows: Candidate sequence generation: On the specified site set L, a candidate sequence set is generated through mutation and recombination operations ; Acquisition function evaluation: For each candidate sequence , calculate the acquisition function value ; Sequence selection and evaluation: Select the sequence with the maximum acquisition function value and perform a true binding affinity evaluation on it to obtain , and ; Surrogate model update: Add new data points to the dataset D and update the surrogate model ; Convergence judgment: If the maximum number of iterations is reached or the preset convergence condition is satisfied, the optimization is stopped.
9. Use of the system according to any one of claims 1 to 7 in the design and optimization of a D-type polypeptide conjugate of PDL1, characterized in that: The amino acid sequence of the D-type polypeptide conjugate is shown in SEQ ID NO.
1.
10. Use of the system according to any one of claims 1 to 7 in the design and optimization of D-type polypeptide conjugates of MDM2, characterized in that: The amino acid sequence of the D-type polypeptide conjugate is shown in any one of SEQ ID NO.3 - 6.
Citation Information
Cited By
Functional nucleic acid de novo design framework and system based on nucleic acid language model
CN121215034A