Diffusion model enhanced evolution-based antibody drug molecule design optimization method

Abstract

The invention discloses an antibody drug molecule design optimization method based on diffusion model enhanced evolution. The method comprises the following steps: 1, constructing a data set for antibody drug molecule design optimization; 2, converting an amino acid sequence into a probability feature vector by using a position weight matrix, and providing a matching input representation mode for the model; 3, the diffusion model learns data distribution through forward noise addition and reverse noise reduction processes, an unclassified guidance mechanism is introduced in a diffusion model generation process, local guidance optimization is performed on feature vector distribution of candidate antibody drug molecules by using a local search strategy, the most potential antibody drug molecules are selected through nearest neighbor sorting, and a data set is updated; and 4, splitting the population into evolutionary antibody drug molecules and model antibody drug molecules according to a dynamic proportion in each iteration, respectively generating offspring populations, and carrying out environment selection. According to the invention, a wider chemical space and a local optimization area of the antibody drug molecule can be explored, so that the antibody drug molecule with higher potential and more effectiveness can be obtained.

Description

Technical Field

[0001] This invention pertains to expensive black-box problems in the field of target optimization, and particularly relates to an antibody drug molecule design optimization method based on diffusion model-enhanced evolution. Background Technology

[0002] In the field of biopharmaceutical development, the industry is known for its long development cycles, high costs, and low success rates. The discovery of lead compounds, as the starting point for innovative drug development, requires novel structures, excellent biological activity, and good druggability. Molecular optimization of lead compounds is a crucial step in determining whether a candidate drug can successfully enter clinical trials. This step involves systematically enhancing activity, reducing toxicity, and optimizing pharmacokinetic properties and chemical stability while maintaining the core pharmacodynamic framework.

[0003] Currently, methods for solving complex black-box optimization problems mainly include surrogate model-based algorithms, multi-fidelity optimization methods, and parallel computing techniques. Multi-fidelity optimization methods improve optimization efficiency by combining objective function evaluation results of different accuracies; however, errors in low-precision models may interfere with the optimization direction, placing high demands on the combination and transformation strategies of multi-fidelity models. Parallel computing techniques significantly reduce the total optimization time by simultaneously evaluating multiple objective function values; however, their parallel efficiency decreases when parallel computing resources are limited or objective function evaluation times are inconsistent, making it difficult to fully leverage their advantages. In solving complex black-box problems based on Bayesian optimization methods, diffusion models, as one of the surrogate models, have shown strong competitiveness. However, the generalization ability of diffusion models is usually limited by their dependence on predefined noise distributions. When facing dynamically changing optimization problems or high-dimensional parameter spaces, the adaptability and accuracy of diffusion models may be challenged; their predefined fixed Gaussian distribution is difficult to quickly approximate the population distribution of the optimization problem, they are sensitive to initial conditions, and they cannot effectively explore the global optimum. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention proposes an antibody drug molecule design optimization method based on diffusion model-enhanced evolution. This method aims to explore a wider chemical space and local optimization regions of antibody drug molecules, thereby obtaining more promising and effective antibody drug molecules. This significantly improves the preclinical development potential of candidate antibody molecules, reduces the failure rate of experimental synthesis and activity screening, and increases the success rate and efficiency of antibody drug development.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a method for designing and optimizing antibody drug molecules based on diffusion model-enhanced evolution, characterized by the following steps: S1. Obtaining antibody drug molecule sets ,in, Indicates the first One antibody molecule, Indicates the total number of molecules; Pick Obtain the first one to be optimized antibody drug molecules and through the RDKit tool Conduct biological attribute assessments and obtain the following results: Combination affinity , Specificity as well as stability Thus, using equation (1) to construct Multi-attribute optimization target : (1) In equation (1), Indicates the search space; S2. Define the maximum number of iterations. Define and initialize the current iteration number as and define the first Antibody drug molecular population ,in, represent The Middle One antibody drug molecule, Represents the size of the population; S3, according to dynamic proportions p Will Divided into the first Evolutionary Antibody Drug Molecular Population With the Generation model antibody drug molecular population ; ∈ ; S4. Using the diffusion model to analyze the first... Generation model antibody drug molecular population Processing yields the first... Generation model antibody drug molecule population ; S5. Using an evolutionary algorithm to process the first... Evolutionary Antibody Drug Molecular Population Perform crossover and mutation operations to generate the first... Generational evolution of antibody drug molecular population ; S6, will and and After merging, we get the first The antibody drug molecule population was merged, and the most fitness-adapted molecule was selected from it. The antibody drug molecule is composed of the first The global optimal antibody drug molecular population And statistics The most adaptable Each antibody drug molecule belongs to Number of antibody drug molecules and Number of antibody drug molecules ; like Then Assign to ; like ,but Assign to ; like Then keep constant; in, It's an increment. ; S7, will Assign to Then, return to step 3 and execute sequentially until... achieve Until then, thus obtaining the first The global optimal antibody drug molecular population .

[0006] The antibody drug molecule design optimization method based on diffusion model-enhanced evolution described in this invention is also characterized in that S4 includes the following steps: S4.1 Calculate the first using formula (1) Generational model population The multi-attribute optimization objective for each antibody drug molecule is used as the fitness value for each antibody drug molecule, and the antibody drug molecule with the maximum fitness value is selected as the first antibody drug molecule. Generational model population The optimal antibody drug molecule ,make fitness value ; S4.2 Utilizing a feature encoder to analyze the model antibody drug molecule population The Middle 1 antibody drug small molecule Encode to build eigenvectors Thus constructing the first Eigenvector matrix ;in, express The Dimensional features; This represents the total number of feature dimensions; and ; S4.3 Construct a diffusion model based on a multilayer perceptron for use in... Perform positive noise addition to obtain Weighted eigenvector distribution ; S4.4, Output the first using the diffusion model In the current time step Fusion prediction Gaussian noise ; S4.5, Using the trained diffusion model to... Perform inverse denoising to obtain the first Distribution of denoised feature vectors ; from = Initially, using equation (6) to... Perform noise reduction until... Until then, thus obtaining the first Distribution of denoised feature vectors And recorded as the number Denoising eigenvector matrix ,in, Indicates the first The generation One denoised feature vector: (6) In equation (8), Indicates the first In the current time step The distribution of denoised feature vectors, Indicates the first In the current time step The distribution of the denoised feature vectors, when t=T, let ; It is the current time step The randomness intensity coefficient, and ; S4.6, Using formula (7) Perform local guidance optimization to obtain the first The generation Optimized denoised feature vectors Thus, the first The optimized denoised feature vector matrix ; (7) In equation (7), It's the learning rate. Indicates in gradient at; Is it a diffusion model? The posterior distribution, Is it a diffusion model? The upper confidence limit UCB score. It is a hyperparameter; S4.7, Selecting from the nearest neighbor sorting method Choose the one with the highest fitness The denoised feature vector corresponding to each antibody drug molecule, and compared with... Choose the one with the lowest fitness. The denoised feature vector corresponding to the i-th antibody drug molecule is replaced to obtain the i-th... Eigenvector matrix ; S4.8, Using a feature decoder to... Transform into the first Generation model antibody drug molecular population ,in, express The Middle A model antibody drug small molecule.

[0007] Furthermore, S4.3 includes the following steps: S4.3.1, Define the current time step as ,and ; S4.3.2, From = Initially, the diffusion model uses equation (2) to... Add current time step The noise, until Reaching the maximum time step Until then, thus obtaining eigenvector distribution : (2) In equation (2), Indicates the first In the current time step eigenvector distribution, It is the current time step The noise figure, At the current time step Standard Gaussian noise, Indicates a standard Gaussian distribution; Indicates the first In time step The eigenvector distribution, when season . Furthermore, S4.4 includes the following steps: S4.4.1, For the current time step Perform sinusoidal position coding to obtain the time embedding amount. ; S4.4.2, will and The input is processed in the diffusion model to obtain the first... In the current time step Unconditional prediction of Gaussian noise Thus, the unconditional loss of the diffusion model can be constructed using equation (3). This is used to train the diffusion model, resulting in a trained unconditional diffusion model. (3) S4.4.3, will , and The input is processed in the diffusion model to obtain the first... In the current time step Conditional prediction of Gaussian noise Thus, the conditional loss of the diffusion model can be constructed using equation (4). This is used to train the diffusion model, resulting in a trained conditional diffusion model: (4) S4.4.4. The trained diffusion model is used to obtain the first... In the current time step Optimal fusion prediction of Gaussian noise : (5) In equation (5), For guiding strength parameters; The first value represents the output of the unconditional diffusion model after training. In the current time step The optimal unconditional prediction of Gaussian noise; The first output of the conditional diffusion model after training represents the... In the current time step The optimal conditional prediction of Gaussian noise.

[0008] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program that supports the processor in executing the antibody drug molecule design optimization method, and the processor is configured to execute the program stored in the memory.

[0009] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program, when run by a processor, executes the steps of the antibody drug molecule design optimization method.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention designs a local search-guided strategy to direct generated antibody drug molecules towards regions with high druggability. After each generation of candidate molecule distributions in the diffusion model, antibody sequences in high-affinity, low-immunogenicity regions are considered dominant molecules, guiding molecules in low-potential regions towards high-activity regions. This combined approach significantly reduces rational bias in traditional computational design and avoids early R&D failures caused by over-focusing on a single metric and neglecting developability in antibody optimization. The feature vectors after local search guidance are more concentrated in high-affinity regions, and the optimized antibody molecules exhibit higher target specificity and druggability in subsequent in vitro binding experiments and in vivo efficacy evaluations. This provides more efficient directional guidance for the precise optimization of antibody drug molecules.

[0011] 2. This invention introduces an adaptive guided controller during antibody molecule generation to achieve a balanced optimization of sequence diversity and candidate molecule quality. The unconditional diffusion model is responsible for broadly exploring the antibody sequence space, while the conditional diffusion model optimizes the quality and functional relevance of the generated molecules by incorporating multi-dimensional conditional information such as affinity, stability, and immunogenicity. This dual mechanism not only expands the search range for rare and high-potential antibody sequences but also ensures synergistic optimization of the generated antibody molecules in terms of complementarity-determining region chemical structure and antigen-binding biological activity, thus better adapting to the complex needs of multi-objective trade-offs in antibody drug molecule design.

[0012] 3. This invention merges candidate antibody molecules generated by a diffusion model, antibody sequences optimized by traditional evolutionary algorithms, and a parent elite population into a single environment selection stage. This fusion mechanism allows data-driven molecule generation and physically-based evolutionary computation to compete and complement each other, ensuring the continuous evolution of the antibody population in terms of sequence diversity and spatial conformational adaptability. The final retention ratio of each source is determined by an adaptive dynamic ratio. This mechanism continuously supplements the underexplored antibody structural space, providing usable diversity and potential for continued exploration in subsequent iterations. This mechanism is particularly suitable for high-dimensional sequence-structure-function space and multi-objective dynamic equilibrium problems in antibody drug molecule optimization, significantly reducing wet screening costs and improving the success rate of candidate antibodies entering CMC development and preclinical studies. Attached Figure Description

[0013] Figure 1 This is a diagram illustrating the process of model training and solution generation in this invention. Figure 2 This is a diagram illustrating the environmental selection and dynamic scaling process of this invention. Figure 3 This is a framework diagram of the method of the present invention. Detailed Implementation

[0014] In this embodiment, to overcome the challenges of high costs for wet experimental validation, lengthy computational simulations, and difficulty in achieving a balance among multiple dimensions such as affinity, specificity, immunogenicity, and developability in existing antibody drug molecule design, a diffusion model-enhanced evolutionary antibody drug molecule design optimization method is proposed. This method learns the distribution patterns of natural antibody sequences and active conformations through a diffusion model, intelligently explores candidate molecules with both high target affinity and low adverse effects in the amino acid sequence space, and uses an evolutionary framework to perform multi-objective synergistic optimization of druggability parameters, aiming to significantly improve the preclinical development potential of candidate antibody molecules and reduce the failure rate of experimental synthesis and activity screening. By combining model-driven molecular generation with directed evolution, the computational and experimental costs of molecular dynamics evaluation and in vitro functional validation are significantly reduced, thereby accelerating the development process from target discovery to preclinical candidate drugs and improving the success rate and efficiency of antibody drug development. Specifically, referring to... Figure 3 As shown, the method is performed according to the following steps: S1. Obtaining antibody drug molecule sets ,in, Indicates the first One antibody molecule, Indicates the total number of molecules; Pick Obtain the first one to be optimized antibody drug molecules and through the RDKit tool Conduct biological attribute assessments and obtain the following results: Combination affinity , Specificity as well as stability Thus, using equation (1) to construct Multi-attribute optimization target : (1) In equation (1), This represents the search space.

[0015] S2. Define the maximum number of iterations. Define and initialize the current iteration number as and define the first Antibody drug molecular population ,in, represent The Middle One antibody drug molecule, It represents the size of the population.

[0016] S3, according to dynamic proportions p Will Divided into the first Evolutionary Antibody Drug Molecular Population With the Generation model antibody drug molecular population ; ∈ .

[0017] S4. Using the diffusion model to analyze the first... Generation model antibody drug molecular population Processing yields the first... Generation model antibody drug molecule population .

[0018] S4.1 Calculate the first using formula (1) Generational model population The multi-attribute optimization objective for each antibody drug molecule is used as the fitness value for each antibody drug molecule, and the antibody drug molecule with the maximum fitness value is selected as the first antibody drug molecule. Generational model population The optimal antibody drug molecule ,make fitness value ; S4.2 Utilizing a feature encoder to analyze the model antibody drug molecule population The Middle 1 antibody drug small molecule Encode to build eigenvectors Thus constructing the first Eigenvector matrix ;in, express The Dimensional features; This represents the total number of feature dimensions; and .

[0019] S4.3 Construct a diffusion model based on a multilayer perceptron for use in... Perform positive noise addition to obtain Weighted eigenvector distribution ; S4.3.1, Define the current time step as ,and .

[0020] S4.3.2, From = Initially, the diffusion model uses equation (2) to... Add current time step The noise, until Reaching the maximum time step Until then, thus obtaining eigenvector distribution : (2) In equation (2), Indicates the first In the current time step eigenvector distribution, It is the current time step The noise figure, At the current time step Standard Gaussian noise, Indicates a standard Gaussian distribution; Indicates the first In time step The eigenvector distribution, when season .

[0021] S4.4, Output the first using the diffusion model In the current time step Fusion prediction Gaussian noise ; S4.4.1, For the current time step Perform sinusoidal position coding to obtain the time embedding amount. ; For the current time step Perform sinusoidal position coding to obtain the time embedding amount. ; , in It is the embedding dimension, a hyperparameter whose value matches the dimension of the selected neural network. It is a component subscript, numbered as follows Each pair For a sine-cosine combination of different frequencies, the frequencies are determined by... Decision. The sum can be expressed as S4.4.2, will and The input is processed in the diffusion model to obtain the first... In the current time step Unconditional prediction of Gaussian noise Thus, the unconditional loss of the diffusion model can be constructed using equation (3). This is used to train the diffusion model, resulting in a trained unconditional diffusion model, as shown in the reference. Figure 1 As shown: (3) S4.4.3, will , and The input is processed in the diffusion model to obtain the first... In the current time step Conditional prediction of Gaussian noise Thus, the conditional loss of the diffusion model can be constructed using equation (4). This is used to train the diffusion model, resulting in a trained conditional diffusion model: (4).

[0022] S4.4.4. The trained diffusion model is used to obtain the first... In the current time step Optimal fusion prediction of Gaussian noise : (5) In equation (5), For guiding strength parameters; The first value represents the output of the unconditional diffusion model after training. In the current time step The optimal unconditional prediction of Gaussian noise; The first output of the conditional diffusion model after training represents the... In the current time step The optimal conditional prediction of Gaussian noise.

[0023] Guiding strength parameters At each time step Dynamic optimization is performed in each backsampling step, based on the current... Calculate the predicted value of the proxy model And update via gradient ascent : in, The learning rate is set to 0.01 by default. It is an assignment operator; S4.5, Using the trained diffusion model to... Perform inverse denoising to obtain the first Distribution of denoised feature vectors ; from = Initially, using equation (6) to... Perform noise reduction until... Until then, thus obtaining the first Distribution of denoised feature vectors And recorded as the number Denoising eigenvector matrix ,in, Indicates the first The generation One denoised feature vector: (6) In equation (8), Indicates the first In the current time step The distribution of denoised feature vectors, Indicates the first In the current time step The distribution of the denoised feature vectors, when t=T, let ; It is the current time step The randomness intensity coefficient, and .

[0024] S4.6, Using formula (7) Perform local guidance optimization to obtain the first The generation Optimized denoised feature vectors Thus, the first The optimized denoised feature vector matrix ; (7) In equation (7), It's the learning rate. Indicates in gradient at; Is it a diffusion model? The posterior distribution is used to maintain a reasonable distribution region, and probability flow ODE estimation is employed. The gradient; Is it a diffusion model? The upper confidence limit UCB score. : It is the mean of the posterior distribution. It is the variance of the posterior distribution; It is a hyperparameter.

[0025] Repeat the gradient ascent steps described above. Next, optimize each step gradually. The final optimized collection is obtained. .

[0026] S4.7, Selecting from the nearest neighbor sorting method Choose the one with the highest fitness The denoised feature vector corresponding to each antibody drug molecule, and compared with... Choose the one with the lowest fitness. The denoised feature vector corresponding to the i-th antibody drug molecule is replaced to obtain the i-th... Eigenvector matrix .

[0027] S4.8, Using a feature decoder to... Transform into the first Generation model antibody drug molecular population ,in, express The Middle A model antibody drug small molecule.

[0028] S5. Using an evolutionary algorithm to process the first... Evolutionary Antibody Drug Molecular Population Perform crossover and mutation operations to generate the first... Generational evolution of antibody drug molecular population .

[0029] S6, Reference Figure 2 As shown, and and After merging, we get the first The antibody drug molecule population was merged, and the most fitness-adapted molecule was selected from it. The antibody drug molecule is composed of the first The global optimal antibody drug molecular population And statistics The most adaptable Each antibody drug molecule belongs to Number of antibody drug molecules and Number of antibody drug molecules ; like Then Assign to ; like ,but Assign to ; like Then keep constant; in, It's an increment. .

[0030] S7, will Assign to Then, return to step 3 and execute sequentially until... achieve Until then, thus obtaining the first The global optimal antibody drug molecular population .

[0031] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0032] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A diffusion model-based enhanced evolutionary method for antibody drug molecule design optimization, characterized in that, Includes the following steps: S1. Obtaining antibody drug molecule sets ,in, Indicates the first One antibody molecule, Indicates the total number of molecules; Pick Obtain the first one to be optimized antibody drug molecules and through the RDKit tool Conduct biological attribute assessments and obtain the following results: Combination affinity , Specificity as well as stability Thus, using equation (1) to construct Multi-attribute optimization target : (1) In equation (1), Indicates the search space; S2. Define the maximum number of iterations. Define and initialize the current iteration number as and define the first Antibody drug molecular population ,in, represent The Middle One antibody drug molecule, Represents the size of the population; S3, according to dynamic proportions p Will Divided into the first Evolutionary Antibody Drug Molecular Population With the Generation model antibody drug molecular population ; ∈ ; S4. Using the diffusion model to analyze the first... Generation model antibody drug molecular population Processing yields the first... Generation model antibody drug molecule population ; S5. Using an evolutionary algorithm to process the first... Evolutionary Antibody Drug Molecular Population Perform crossover and mutation operations to generate the first... Generational evolution of antibody drug molecular population ; S6, will and and After merging, we get the first The antibody drug molecule population was merged, and the most fitness-adapted molecule was selected from it. The antibody drug molecule is composed of the first The global optimal antibody drug molecular population And statistics The most adaptable Each antibody drug molecule belongs to Number of antibody drug molecules and Number of antibody drug molecules ; like Then Assign to ; like ,but Assign to ; like Then keep constant; in, It's an increment. ; S7, will Assign to Then, return to step 3 and execute sequentially until... achieve Until then, thus obtaining the first The global optimal antibody drug molecular population .

2. The antibody drug molecule design optimization method based on diffusion model-enhanced evolution as described in claim 1, characterized in that, S4 includes the following steps: S4.1 Calculate the first using formula (1) Generational model population The multi-attribute optimization objective for each antibody drug molecule is used as the fitness value for each antibody drug molecule, and the antibody drug molecule with the maximum fitness value is selected as the first antibody drug molecule. Generational model population The optimal antibody drug molecule ,make fitness value ; S4.2 Utilizing a feature encoder to analyze the model antibody drug molecule population The Middle 1 antibody drug small molecule Encode to build eigenvectors Thus constructing the first Eigenvector matrix ;in, express The Dimensional features; This represents the total number of feature dimensions; and ; S4.3 Construct a diffusion model based on a multilayer perceptron for use in... Perform positive noise addition to obtain Weighted eigenvector distribution ; S4.4, Output the first using the diffusion model In the current time step Fusion prediction Gaussian noise ; S4.5, Using the trained diffusion model to... Perform inverse denoising to obtain the first Distribution of denoised feature vectors ; from = Initially, using equation (6) to... Perform noise reduction until... Until then, thus obtaining the first Distribution of denoised feature vectors And recorded as the number Denoising eigenvector matrix ,in, Indicates the first The generation One denoised feature vector: (6) In equation (8), Indicates the first In the current time step The distribution of denoised feature vectors, Indicates the first In the current time step The distribution of the denoised feature vectors, when t=T, let ; It is the current time step The randomness intensity coefficient, and ; S4.6, Using formula (7) Perform local guidance optimization to obtain the first The generation Optimized denoised feature vectors Thus, the first The optimized denoised feature vector matrix ; (7) In equation (7), It's the learning rate. Indicates in gradient at; Is it a diffusion model? The posterior distribution, Is it a diffusion model? The upper confidence limit UCB score. It is a hyperparameter; S4.7, Selecting from the nearest neighbor sorting method Choose the one with the highest fitness The denoised feature vector corresponding to each antibody drug molecule, and compared with... Choose the one with the lowest fitness. The denoised feature vector corresponding to the i-th antibody drug molecule is replaced to obtain the i-th... Eigenvector matrix ; S4.8, Using a feature decoder to... Transform into the first Generation model antibody drug molecular population ,in, express The Middle A model antibody drug small molecule.

3. The antibody drug molecule design optimization method based on diffusion model-enhanced evolution as described in claim 2, characterized in that, S4.3 includes the following steps: S4.3.1, Define the current time step as ,and ; S4.3.2, From = Initially, the diffusion model uses equation (2) to... Add current time step The noise, until Reaching the maximum time step Until then, thus obtaining eigenvector distribution : (2) In equation (2), Indicates the first In the current time step eigenvector distribution, It is the current time step The noise figure, At the current time step Standard Gaussian noise, Indicates a standard Gaussian distribution; Indicates the first In time step The eigenvector distribution, when season .

4. The antibody drug molecule design optimization method based on diffusion model-enhanced evolution as described in claim 3, characterized in that, S4.4 includes the following steps: S4.4.1, For the current time step Perform sinusoidal position coding to obtain the time embedding amount. ; S4.4.2, will and The input is processed in the diffusion model to obtain the first... In the current time step Unconditional prediction of Gaussian noise Thus, the unconditional loss of the diffusion model can be constructed using equation (3). This is used to train the diffusion model, resulting in a trained unconditional diffusion model. (3) S4.4.3, will , and The input is processed in the diffusion model to obtain the first... In the current time step Conditional prediction of Gaussian noise Thus, the conditional loss of the diffusion model can be constructed using equation (4). This is used to train the diffusion model, resulting in a trained conditional diffusion model: (4) S4.4.

4. The trained diffusion model is used to obtain the first... In the current time step Optimal fusion prediction of Gaussian noise : (5) In equation (5), For guiding strength parameters; The first value represents the output of the unconditional diffusion model after training. In the current time step The optimal unconditional prediction of Gaussian noise; The first output of the conditional diffusion model after training represents the... In the current time step The optimal conditional prediction of Gaussian noise.

5. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing the antibody drug molecule design optimization method according to any one of claims 1-4, and the processor is configured to execute the programs stored in the memory.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by the processor, performs the steps of the antibody drug molecule design optimization method according to any one of claims 1-4.

Images