A multi-objective drug molecule generation and optimization method and related apparatus

Abstract

The application provides a multi-target drug molecule generation and optimization method and related device, and relates to the technical field of artificial intelligence and computer-aided drug design. The method comprises the following steps: obtaining a drug molecule database, extracting an initial molecular population, and converting the initial molecular population into a one-dimensional sequence list; performing multi-dimensional fitness evaluation on the sequenced initial population to obtain multi-dimensional independent scores and aggregate a scalar reference total score to form a current molecular population; screening a parent set for mating based on the scalar reference total score; performing cross and mutation on the parents to generate offspring, and merging the offspring with the parents to form a hybrid molecular population; after re-evaluating the multi-dimensional fitness of the hybrid population, screening the hybrid population based on the multi-dimensional independent scores to obtain a new generation of molecular population; and repeating the iteration until a preset number of iterations is reached and outputting. The application realizes efficient multi-target optimization of drug molecules through the iterative closed-loop mechanism of multi-dimensional fitness evaluation and aggregation, cross and mutation, re-evaluation and screening, and effectively solves the problems of scalar collapse, population prematureness and imbalance of chemical rationality.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and computer-aided drug design technology, specifically to a method and related apparatus for generating and optimizing multi-target drug molecules. Background Technology

[0002] New drug development is a time-consuming, costly, and high-risk systematic project. Lead compound optimization is a key step in the early stages of development. It mainly involves optimizing the structure of lead compounds and precursor compounds to improve their physicochemical properties, druggability, and synthetic accessibility while retaining or enhancing their target affinity, thereby screening out candidate drugs with clinical potential.

[0003] With the development of AI-assisted drug discovery technology, existing drug molecule design methods still face many limitations: multiple property targets may compete with each other, and simple weighting can easily lead to a bias in the search towards some targets; selection pressure and diversity maintenance in genetic search need to be balanced; random graph structure recombination or structure legality checks alone cannot guarantee that the generated connection mode is consistent with known reaction rules.

[0004] Therefore, there is a need for a multi-objective drug molecule generation and optimization method that can simultaneously utilize multidimensional property evaluation and reaction knowledge, and configure environmental selection strategies based on target space characteristics. Summary of the Invention

[0005] This application addresses the issues of insufficient syntheticity in multi-objective property trade-offs, structure exploration, and computational synthesis of generated molecules by providing a method and related apparatus for multi-objective drug molecule generation and optimization.

[0006] To achieve the above objectives, this application adopts the following technical solution: In a first aspect, this application provides a method for generating and optimizing multi-target drug molecules, comprising the following steps: S1. Obtain a drug molecule database corresponding to the target point, extract a preset number of drug molecules from the drug molecule database as an initial molecular population, and convert the small molecule compounds in the initial molecular population into a one-dimensional molecular sequence representation to obtain a serialized initial population. S2, perform multi-dimensional fitness evaluation on the initial serialized population to obtain multi-dimensional independent scores for each drug molecule, and aggregate the multi-dimensional independent scores to obtain a scalar reference total score for each drug molecule. The current molecular population is formed by having each drug molecule correspond to its respective scalar reference total score. S3, Based on the scalar reference total score, a set of mating parents is selected from the current molecular population; S4, perform crossover and mutation operations on the parent drug molecules in the mating parent set to generate offspring drug molecules, and merge the parent drug molecules with the offspring drug molecules to form a mixed molecular population; S5, perform a multi-dimensional fitness assessment on the mixed molecular population again to obtain the multi-dimensional independent score corresponding to each drug molecule in the mixed molecular population; S6. Based on the multidimensional independent scores, drug molecules are screened from the mixed molecular population to obtain a new generation of molecular populations. S7. Take the new generation of molecular population obtained in step S6 as the current molecular population for the next iteration, and repeat steps S3 to S6 until the preset number of iterations is reached, and output the optimized target molecular population.

[0007] Furthermore, the process of step S2 specifically includes: The original attribute scores of each drug molecule under multiple preset target dimensions are calculated using a corresponding attribute scorer. Based on the preset optimization direction of each evaluation index, the original attribute scores that do not have uniform monotonicity are subjected to monotonicity transformation processing to obtain the uniform monotonicity score of each drug molecule. The monotonicity score of each drug molecule is input into the bimodal fitness conversion mechanism. Based on the optimization task stage or preset instructions, the continuous mode or discrete mode is selected for mapping conversion to form the multidimensional independent score of each drug molecule. The continuous mode retains the continuous floating-point gradient of the monotonic uniform score and performs normalization processing, while the discrete mode sets a corresponding evaluation hard threshold for each preset target dimension and forcibly converts the monotonic uniform score into a discrete binary score of 0 or 1. The multidimensional independent scores corresponding to each drug molecule in each preset target dimension are linearly weighted or summed to obtain the scalar reference total score with a single value.

[0008] Furthermore, step S3 specifically includes: Each time, a preset number of drug molecules are randomly selected from the current molecular population as candidates. The scalar reference total score of each candidate is compared. Each time, the candidate with the highest scalar reference total score is selected as a mating parent. This process is repeated until the set of mating parents reaches the preset population size.

[0009] Furthermore, in step S4, a crossover operation is performed on the parent drug molecules in the mating parent set, specifically including: The one-dimensional molecular sequence representation of the paired parent drug molecule is input as the reactant sequence into an attention-based molecular converter model. The molecular converter model predicts the product molecule sequence based on the reaction mapping relationship learned in the pre-training stage, and the predicted product that passes the structural legality check is used as the candidate progeny drug molecule.

[0010] Furthermore, step S4 involves performing mutation operations on the parent drug molecules in the mating parent set, specifically including: According to the preset mutation probability, the molecular structure of drug molecules is applied to a preset chemical reaction template composed of reaction rules defined by SMARTS, so as to achieve local fine-tuning of the molecular structure.

[0011] Furthermore, step S6 specifically includes: Based on the multidimensional independent scores, non-dominated sorting is performed on the mixed molecular population to construct multiple Pareto levels; The number of preset target dimensions in step S2 is used as the number of optimization targets and compared with a preset dimension threshold. For drug molecules within the same Pareto level, if the number of the preset target dimensions is less than or equal to the dimension threshold, a diversity maintenance strategy based on a combination of crowding distance and random exploration is executed; if the number of the preset target dimensions is greater than the dimension threshold, a niche diversity assessment strategy based on reference point association is executed; and the next generation of molecular populations is screened based on the executed strategy.

[0012] Furthermore, the specific process of implementing the diversity maintenance strategy based on a fusion of crowding distance and random exploration is as follows: Combining the non-dominated ranking order and crowding distance of drug molecules, elite drug molecules at the forefront are preferentially screened according to the first preset retention quantity; Random sampling is performed from the remaining non-elite drug molecules in the mixed molecular population, according to a second preset retention quantity; The selected elite drug molecules are combined with randomly sampled drug molecules to form the new generation of molecular population. The specific process for implementing the reference point-based microhabitat diversity assessment strategy is as follows: The priority order of drug molecules is determined based on their non-dominated ordering. Calculate the correlation distance between each drug molecule and a pre-defined reference point in the target space; Based on the correlation distance and niche count results, drug molecules within the same Pareto level are screened; A predetermined number of drug molecules are extracted from the mixed molecular population to form the new generation molecular population.

[0013] Furthermore, the attribute scorer includes at least one of a machine learning-based target activity predictor, a heuristic calculator based on physicochemical rules of thumb, and a statistical scorer based on molecular fragments and structural complexity. The monotonicity transformation process includes: applying a linear extreme value mapping and inversion formula to the original score for the one-sided extreme value constraint attribute to convert the original score into a normalized score where the larger the value, the better the performance; and introducing a bell-shaped nonlinear normalization formula based on the Gaussian distribution function for the two-sided interval constraint attribute. When the original score of the drug molecule hits the preset optimal target bias, the score is full; if it deviates from the optimal target bias, the score smoothly decays towards zero.

[0014] Secondly, this application provides a multi-target drug molecule generation and optimization system, comprising: The database acquisition module is used to acquire a database of drug molecules corresponding to the target. An initialization module is used to extract a preset number of drug molecules from the drug molecule database as an initial molecular population, and to convert the small molecule compounds in the initial molecular population into a one-dimensional molecular sequence representation to obtain a serialized initial population. The evaluation aggregation module is used to perform multi-dimensional fitness evaluation on the initial serialized population, obtain multi-dimensional independent scores for each drug molecule, and aggregate the multi-dimensional independent scores to obtain a scalar reference total score for each drug molecule. The current molecular population is constructed in such a way that each drug molecule corresponds to its respective scalar reference total score. The mating screening module is used to screen out a set of mating parents from the current molecular population based on the scalar reference total score; The crossover and mutation module is used to perform crossover and mutation operations on the parent drug molecules in the mating parent set to generate offspring drug molecules, and to merge the parent drug molecules with the offspring drug molecules to form a mixed molecular population; The re-evaluation module is used to re-evaluate the multi-dimensional fitness of the mixed molecular population to obtain the multi-dimensional independent score corresponding to each drug molecule in the mixed molecular population. A screening module is used to screen drug molecules from the mixed molecular population based on the multidimensional independent scores to obtain a new generation of molecular populations. The loop control module is used to repeatedly trigger the mating screening module, the crossover mutation module, the re-evaluation module, and the screening module to perform operations until a preset number of iterations is reached, and output the optimized target molecular population.

[0015] Thirdly, this application provides a computer device, including: a processor and a computer-readable storage medium; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the method for generating and optimizing multi-target drug molecules.

[0016] Compared with the prior art, this application has the following beneficial effects: This application provides a method for multi-objective drug molecule generation and optimization. This method, while obtaining a scalar reference total score for parental selection through aggregation, retains the multidimensional independent scores of each molecule and performs non-dominated environmental selection based on these multidimensional independent scores in a combined parent and offspring population, thus avoiding the direct determination of all survival selection by a single aggregation score. An iterative closed loop is constructed through multidimensional scoring, parental selection, response-sensing crossover, local mutation, joint re-evaluation, and environmental selection, enabling the population to gradually form compromise solutions among multiple properties. In the four-objective discrete embodiment of this application, the perfect molecule set contains 3066 non-repeating Bemis-Murcko skeletons.

[0017] Furthermore, this application uses continuous scoring and discrete threshold scoring as two configurable fitness forms. Continuous scoring preserves property differences, while discrete scoring converts each objective to 0 or 1 based on a preset threshold to alter parental selection pressure. In a single experiment, the novelty of the three-objective continuous and discrete setups were 61.2% and 86.0%, respectively, with 1378 and 767 backbones, respectively; the novelty of the four-objective continuous and discrete setups were 48.0% and 87.0%, respectively, with 1618 and 3066 backbones, respectively. These results indicate that discrete scoring improves novelty in this embodiment, but its impact on backbone coverage varies with the task. Parental selection employs a binary tournament mechanism, using a scalar reference total score for local comparison. In the crossover stage, parental molecular sequences are input as reactants into a pre-trained molecular converter, and predicted products are used as candidate progeny. In the six-target cross-ablation of 300 candidate molecules in each group, the computational synthesizability corresponding to GB-GA graph structure cross and molecular converter cross were 75.0% and 91.3%, respectively, indicating that reaction-sensing cross is associated with a higher computational synthesizability under this experimental setting.

[0018] Furthermore, this application performs non-dominated ordination in a combined parent and offspring population and can select different diversity maintenance strategies based on the number of targets or a preset configuration: screening based on crowding and random exploration is used in lower-dimensional target spaces, while screening based on reference direction association and niche counting is used in higher-dimensional target spaces. In comparisons of 300 candidate molecules per group, NSGA-II and NSGA-III showed different preferences in three-, four-, and six-target tasks. In the six-target task, their computational synthesizability was 75.6% and 91.3%, respectively; novelty was 32.6% and 35.7%, respectively; diversity was 0.7267 and 0.7200, respectively; and the number of skeletons was 144 and 139, respectively. These results indicate that the reference direction strategy improves the computational synthesizability under this six-target setting, but diversity and the number of skeletons do not increase synchronously. Therefore, the environmental selection strategy should be configured in conjunction with the task dimension and target preference. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a system flowchart of a multi-target drug molecule generation and optimization method provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] See Figure 1 This application provides a method for generating and optimizing multi-target drug molecules, comprising the following steps: S1. Obtain a drug molecule database corresponding to the target point, extract a preset number of drug molecules from the drug molecule database as an initial molecular population, and convert the small molecule compounds in the initial molecular population into a one-dimensional molecular sequence representation to obtain a serialized initial population. In a more specific embodiment provided in this application, step S1 is performed as follows: Identify the target. The target can be a protein, enzyme, or receptor associated with a specific disease, such as the kinase targets JNK3 and GSK3β.

[0023] Obtain a database of drug molecules corresponding to the target. This database can be a publicly available compound database, such as the ChEMBL, ZINC, or PubChem databases, or a proprietary database built for a specific target. For example, in the dual-target optimization of JNK3 and GSK3β, drug molecules known to be active against JNK3 or GSK3β can be screened from the ChEMBL database to construct an initial database.

[0024] A predetermined number of drug molecules are extracted from a drug molecule database as an initial molecular population. This predetermined number can be set based on computational resources and the complexity of the optimization task. For example, it can be set to 100, 200, or 500 molecules. In some implementations, the initial molecular population size is set to 100 molecules. The extraction method can employ random sampling or maximum differential sampling based on molecular diversity to ensure that the initial population has broad structural coverage.

[0025] The small molecule compounds in the initial molecular population are converted into one-dimensional molecular sequence representations to obtain the serialized initial population. The one-dimensional molecular sequence representation can be either the SMILES string or the SELFIES string.

[0026] Through the above transformation process, each drug molecule is converted into a string sequence, and all sequences together constitute the initial population of serialization.

[0027] S2, perform multi-dimensional fitness evaluation on the initial serialized population to obtain multi-dimensional independent scores for each drug molecule, and aggregate the multi-dimensional independent scores to obtain a scalar reference total score for each drug molecule. The current molecular population is formed by having each drug molecule correspond to its respective scalar reference total score. To comprehensively evaluate the overall drug potential of the generated candidate drug molecules, the system pre-sets multi-dimensional evaluation indicators. These indicators include, but are not limited to: target affinity, quantitative drug-likeness, synthetic accessibility, topological polar surface area, and lipid-water partition coefficient.

[0028] The target activity score characterizes the model prediction probability of a candidate small molecule for a specific target (such as DRD2, JNK3, or GSK3β), with a value ranging from 0 to 1. A higher value indicates a higher probability that the model predicts that the molecule satisfies the corresponding activity tag. Therefore, the target activities reported in this embodiment are model prediction scores, not experimentally determined activities.

[0029] Quantitative Estimate of Drug-likeness (QED) is used to comprehensively measure the similarity of a molecule to a known oral drug in multiple physicochemical dimensions, such as molecular weight, number of hydrogen bond donors and acceptors, number of rotatable bonds, and number of aromatic rings. The value ranges from 0 to 1, with a higher value indicating better drug-likeness.

[0030] The Synthetic Accessibility score (SA score) is a heuristic estimate of the difficulty of synthesis based on the contribution of molecular fragments and structural complexity. The original value is usually between 1 and 10, and the lower the value, the lower the estimated synthesis difficulty.

[0031] Topological polar surface area (TPSA) is often used to characterize molecular polarity and is related to properties such as membrane permeability and oral absorption; while the lipid-water partition coefficient (LogP) reflects the hydrophobicity and lipid solubility of molecules.

[0032] In this embodiment, TPSA and LogP are converted into interval-preference continuous scores to guide candidate molecules toward a preset range of preferred physicochemical properties.

[0033] In a more specific embodiment provided in this application, step S2 is performed as follows: S2.1 Calculate the original attribute scores of each drug molecule under multiple preset target dimensions using the corresponding attribute scorer; Machine learning-based target activity predictor: Load pre-trained model weights, convert the one-dimensional sequence of drug molecules into Morgan fingerprints, and output the predicted probability of the molecule belonging to the corresponding activity category through a random forest classifier.

[0034] Heuristic calculators based on physicochemical rules of thumb: relying on underlying algorithms of cheminformatics (such as the RDKit open-source library), they directly analyze the two-dimensional topological structure of drug molecules and calculate their quantitative drug-like properties and physicochemical parameters through physicochemical equations.

[0035] A statistical scorer based on molecular fragments and structural complexity: The SA score is calculated based on the fragment contribution and structural complexity of candidate molecules, and is used to characterize the synthesis difficulty estimated by the model.

[0036] S2.2, Based on the preset optimization direction of each evaluation index, perform monotonicity transformation on the original attribute scores that do not have a unified monotonicity in terms of the original numerical distribution and performance quality, so that the scores of all target dimensions after processing satisfy the consistency that the larger the value, the better the performance, and obtain a unified monotonicity score for each drug molecule. To unify the optimization direction of each objective, the SA score is converted into a maximization form where a larger value indicates better performance:

[0037] Here, SA represents the original synthetic accessibility score, and SA_norm represents the normalized score. This transformation maps the original range of 1 to 10 to 0 to 1, and limits the minimum score to 0 through max(0,·); the lower the original SA score, the higher the SA_norm.

[0038] For TPSA and LogP, a Gaussian continuous rating with a preset preference center as the peak value is adopted, and its general normalization formula is as follows:

[0039] Where x represents the physicochemical property value to be evaluated, μ is the center of preference, and σ is the scaling parameter. In this embodiment, the center of preference and scaling parameter of TPSA are set to 60 and 30, respectively, and the center of preference and scaling parameter of LogP are set to 3.0 and 1.5, respectively. The specific normalization formula is as follows:

[0040]

[0041] Here, TPSA_norm and LogP_norm represent the normalized scores of TPSA and LogP, respectively. The score is 1 when the original value of the candidate molecule is equal to the corresponding preference center, and the score decreases smoothly according to a Gaussian function when it deviates from the preference center.

[0042] S2.3, the monotonicity score of each drug molecule is input into the bimodal fitness conversion mechanism, and the continuous mode or discrete mode is selected for mapping conversion according to the optimization task stage or preset instructions to form a multidimensional independent score for each drug molecule.

[0043] The continuous mode preserves the continuous floating-point gradient of the monotonicity uniform score and performs normalization to distinguish the subtle property differences of candidate molecules on each target.

[0044] Discrete modalities set evaluation thresholds for each preset target dimension, converting whether a candidate molecule meets the corresponding requirements into 0 or 1. For example, in three-target and four-target tasks, thresholds such as a target activity score of no less than 0.5, a QED of no less than 0.6, and an SA score of no more than 4.0 can be used, respectively. A value of 1 is assigned when the corresponding condition is met, and 0 is assigned otherwise. This modality changes the way scores are distinguished after reaching the threshold, thus creating a different selection pressure than the continuous modality.

[0045] S2.4, perform linear weighting or summation on the multidimensional independent scores of each preset target dimension corresponding to each drug molecule to obtain the scalar reference total score of a single value corresponding to each drug molecule.

[0046] The current molecular population is constructed by assigning a scalar reference total score to each drug molecule. The initial sequenced population is then transformed into the current molecular population carrying the scalar reference total score.

[0047] S3, Based on the scalar reference total score, a set of mating parents is selected from the current molecular population; In a more specific embodiment provided in this application, step S3 is performed as follows: A tournament selection mechanism is used to select mating sires from the current molecular population. Each round of the tournament randomly selects a predetermined number K different candidate molecules from the current population and compares their scalar reference total scores; the same molecule is allowed to be selected again between different rounds. In some implementations, K=2, i.e., a binary tournament selection is used.

[0048] Compare the scalar reference total scores of the candidates in this round and select the one with the higher score as a mating paternal parent. Repeat the tournament selection to form the required number of parental pairs; when two molecules in the same parental pair are identical, select a second parent.

[0049] Repeat the above tournament selection and parent pairing process until a predetermined number of different parent combinations are formed.

[0050] This selection mechanism determines parents based on the relative scores of local candidates. Candidates can participate in the tournament repeatedly in different rounds, and the tournament size K is used to regulate selection pressure. Compared with direct truncation based on the absolute score of the entire population, this mechanism can maintain a certain degree of randomness while preserving the selection opportunities for high-scoring individuals.

[0051] S4, perform crossover and mutation operations on the parent drug molecules in the mating parent set to generate offspring drug molecules, and merge the parent drug molecules with the offspring drug molecules to form a mixed molecular population; In a more specific embodiment provided in this application, step S4 is performed as follows: This step includes two parts: crossover and mutation.

[0052] For the crossover operation: Repeat the binary tournament selection to form the required number of different parent pairs, and sort the SMILES strings of each parent pair in lexicographical order and connect them with the delimiter "." to form the reactant sequence input in the form of "SMILES_A.SMILES_B".

[0053] The molecular converter model employs a Transformer-based pre-trained forward reaction prediction model. The model receives reactant sequences composed of parent molecules and outputs one or more candidate products (SMILES) based on the reactant-to-product mapping learned during pre-training. After structural analysis, validity checks, and duplicate removal, the predicted products are used as candidate progeny in subsequent mutation and scoring steps. This application does not alter the network structure of the pre-trained model.

[0054] Cross operations have the following characteristics: The molecular converter learns the statistical mapping between reactants and products from reaction data, thus constraining the generation of candidate progeny by the learned reaction patterns. This constraint helps reduce some irrational connections that may result from random graph structure splicing. However, the model prediction results still need to be evaluated through structural legitimacy checks and subsequent calculations, and cannot be used to directly conclude that the experimental synthetic route is necessarily feasible.

[0055] For mutation operations: Local structural transformations driven by SMARTS rules are applied to candidate offspring according to a preset mutation probability. In this embodiment, the mutation probability is set to 0.05, and the operations include atom insertion, atom appending, atom deletion, atom type replacement, bond order change, ring bond deletion, and ring structure generation. The mutated products proceed to the subsequent scoring steps after undergoing structural validity and ring structure rationality checks.

[0056] For population merging: The parent drug molecules involved in mating are merged with the offspring drug molecules obtained through crossover and mutation operations to form a mixed molecular population. For example, if the set of parent drug molecules contains 100 molecules and the set of offspring drug molecules also contains 100 molecules, the merged population will have 200 molecules.

[0057] S5, perform a multi-dimensional fitness assessment on the mixed molecular population again to obtain the multi-dimensional independent score corresponding to each drug molecule in the mixed molecular population; In a more specific embodiment provided in this application, step S5 is performed as follows: The mixed molecular population formed in step S4 is reassessed for multidimensional fitness. This assessment process is the same as in step S2.

[0058] The original attribute scores of each drug molecule in the mixed molecular population under multiple preset target dimensions are calculated using corresponding attribute scorers. The attribute scorers used are the same type as those in step S2, including a machine learning-based target affinity predictor, a heuristic calculator based on physicochemical rules of thumb, and a statistical scorer based on historical reaction fragments.

[0059] Based on the preset optimization directions of each evaluation index, monotonicity transformation is performed on the original attribute scores that do not have uniform monotonicity to obtain a uniform monotonicity score for each drug molecule. For unilateral extreme value constraint attributes such as SA, linear extreme value mapping and inversion formulas are applied; for bilateral interval constraint attributes such as TPSA and LogP, bell-shaped nonlinear normalization formulas based on Gaussian distribution functions are applied.

[0060] The monotonicity score of each drug molecule is input into the bimodal fitness conversion mechanism. Based on the optimization task stage or preset instructions, the continuous mode or discrete mode is selected for mapping and conversion, thus forming a multidimensional independent score for each drug molecule.

[0061] Through this reassessment step, each drug molecule in the mixed molecular population received an updated multidimensional independent score.

[0062] S6. Based on the multidimensional independent scores, drug molecules are screened from the mixed molecular population to obtain a new generation of molecular populations. In a more specific embodiment provided in this application, step S6 is performed as follows: This step employs an adaptive Pareto optimization strategy for individual selection.

[0063] S6.1, based on multidimensional independent scores, performs non-dominated ranking on a mixed molecular population to construct multiple Pareto levels. Non-dominated ranking is defined as follows: for any two drug molecules A and B, if A's score is not lower than B's in all preset target dimensions, and is strictly higher than B's in at least one dimension, then A is said to dominate B. By comparing the dominance relationships among all drug molecules in the population, the population can be divided into multiple Pareto levels: the first level is the non-dominated solution set not dominated by any other molecule, the second level is the solution set dominated only by molecules in the first level, and so on.

[0064] The number of preset target dimensions is compared with a preset dimension threshold, or the environment selection strategy is determined based on the task configuration. In some implementations, the dimension threshold is set to 4.

[0065] Select a filtering strategy based on the comparison results or preset configuration: If the number of optimization targets is less than or equal to the dimensionality threshold (low-dimensional optimization stage), a diversity maintenance strategy based on a fusion of crowding distance and random exploration is implemented. The crowding distance is calculated as follows: for drug molecules within the same Pareto level, they are sorted by score in each target dimension. The score difference between two adjacent molecules in that dimension is calculated, and the score differences in each dimension are normalized and summed to obtain the crowding distance of that molecule. The specific screening process includes: prioritizing molecules in higher-level Pareto levels based on their non-dominated ranking; within the same level, combining the non-dominated ranking and crowding distance of drug molecules, elite drug molecules at the top frontier are selected according to a first preset retention quantity; random sampling is performed from the remaining non-elite drug molecules in the mixed molecular population according to a second preset retention quantity; the selected elite drug molecules are merged with the randomly sampled drug molecules to form a new generation of molecular population.

[0066] If the number of optimization targets exceeds the dimensionality threshold (high-dimensional optimization stage), a niche diversity assessment strategy based on reference point association is implemented. The specific process of this strategy includes: determining priority based on the non-dominated ranking of drug molecules; constructing uniform reference directional rays (reference points) on the target hyperplane, for example, using the Das-Dennis method to generate uniformly distributed reference points on the hyperplane in a 6-dimensional target space; calculating the association distance between each drug molecule and a preset reference point in the target space, i.e., calculating the perpendicular distance by connecting the multidimensional independent fractional vector of each molecule to each reference point, and finding the reference point closest to the molecule; screening drug molecules within the same Pareto level based on the association distance and niche count results (i.e., the number of molecules selected for each reference point), prioritizing molecules associated with reference points that have fewer niche counts; and extracting a preset number of drug molecules from the mixed molecular population as a new generation of molecular populations.

[0067] S7. Take the new generation of molecular population obtained in step S6 as the current molecular population for the next iteration, and repeat steps S3 to S6 until the preset number of iterations is reached, and output the optimized target molecular population.

[0068] In a more specific embodiment provided in this application, step S7 is performed as follows: The new generation of molecular population obtained in step S6 is used as the current molecular population for the next iteration, and steps S3 to S6 are repeated. Each iteration includes: tournament selection based on scalar reference total score, crossover operation based on molecular converter and mutation operation based on reaction template, re-evaluation of the mixed molecular population, and individual selection based on adaptive Pareto strategy. Each iteration causes the molecular population to evolve towards the optimization objective.

[0069] The number of iterations can be set according to computing resources and task complexity. In the three-objective, four-objective, and six-objective experiments of this embodiment, the population size is set to 100, the number of iterations is set to 50 generations, and the mutation probability is set to 0.05.

[0070] When the number of iterations reaches a preset value, the iteration stops, and the current molecular population is output as the optimized target molecular population. Each drug molecule in the output target molecular population carries its multidimensional independent score and scalar reference total score.

[0071] In another embodiment provided in this application, a multi-target drug molecule generation and optimization system is provided, which is used to perform the methods described above.

[0072] The system includes: a database acquisition module, an initialization module, an evaluation and aggregation module, a mating and screening module, a crossover and mutation module, a re-evaluation module, a screening module, and a loop control module.

[0073] The database acquisition module is used to acquire drug molecule databases corresponding to the target. The target can be a protein, enzyme, or receptor associated with a specific disease, such as the kinase targets JNK3 and GSK3β. The drug molecule database can be a publicly available compound database, such as the ChEMBL, ZINC, or PubChem databases, or a proprietary database built for a specific target.

[0074] The initialization module extracts a preset number of drug molecules from a drug molecule database as an initial molecular population, and converts the small molecule compounds in the initial molecular population into one-dimensional molecular sequence representations, resulting in a serialized initial population. The preset number can be set according to computational resources and optimization task complexity, for example, 100. The one-dimensional molecular sequence representation can be either the string "SMILES" or the string "SELFIES".

[0075] The evaluation aggregation module performs multi-dimensional fitness evaluation on the initial serialized population, obtaining multi-dimensional independent scores for each drug molecule. These independent scores are then aggregated to obtain a scalar reference total score for each drug molecule, forming the current molecular population. Specifically, the evaluation aggregation module calculates the original attribute scores for each drug molecule across multiple preset target dimensions using a corresponding attribute scorer. Based on the preset optimization direction for each evaluation index, it performs monotonicity transformation on the original attribute scores that lack uniform monotonicity, obtaining a uniform monotonic score for each drug molecule. This uniform monotonic score is then input into a dual-modal fitness transformation mechanism. Depending on the optimization task stage or preset instructions, it selects either a continuous or discrete mode for mapping transformation, thus forming the multi-dimensional independent scores for each drug molecule. For the continuous mode, the continuous floating-point gradient of the uniform monotonic score is preserved and normalized. For the discrete mode, a corresponding hard threshold is set for each preset target dimension, and the uniform monotonic score is forcibly converted into a discrete binary score of 0 or 1. The evaluation aggregation module also performs linear weighting or summation on the multidimensional independent scores of each drug molecule for each preset target dimension to obtain a single numerical scalar reference total score.

[0076] The mating selection module is used to select a set of mating parents from the current molecular population based on the scalar reference total score. The mating selection module adopts a tournament selection mechanism, randomly selecting a preset number of drug molecules as candidates from the current molecular population each time, comparing the scalar reference total scores of each candidate, and selecting the candidate with the highest total score as a mating parent. This process is repeated until the set of mating parents reaches the preset population size.

[0077] The crossover and mutation module performs crossover and mutation operations on parent drug molecules in the mating set to generate daughter drug molecules. It then merges the parent and daughter drug molecules to form a mixed molecular population. Specifically, the crossover operation involves inputting the one-dimensional molecular sequence representations of the paired parent drug molecules into an attention-based molecular converter model. This model performs crossover and fusion of the parent drug molecule structures in the latent space and generates daughter drug molecules based on the chemical reaction rules learned during model pre-training. The molecular converter model is a pre-trained model based on the Transformer architecture. The mutation operation involves applying a preset chemical reaction template, constructed using reaction rules defined by SMARTS, to the molecular structure of the drug molecules according to preset mutation probabilities, achieving local fine-tuning of the molecular structure.

[0078] The re-evaluation module is used to re-evaluate the multi-dimensional fitness of the mixed molecular population, obtaining multi-dimensional independent scores for each drug molecule in the mixed molecular population. The evaluation method used in the re-evaluation module is the same as that used in the evaluation aggregation module.

[0079] The screening module is used to screen drug molecules from mixed molecular populations based on multidimensional independent scores to obtain a new generation of molecular populations. The screening module employs an adaptive Pareto optimization strategy: it performs non-dominated ordination on the mixed molecular population based on multidimensional independent scores to construct multiple Pareto levels. The number of preset target dimensions is compared with a preset dimensionality threshold. For drug molecules within the same Pareto level, if the number of preset target dimensions is less than or equal to the dimensionality threshold, a diversity maintenance strategy combining crowding distance and random exploration is executed; if the number of preset target dimensions is greater than the dimensionality threshold, a niche diversity assessment strategy based on reference point association is executed. Based on the executed strategies, a new generation of molecular populations is screened.

[0080] The loop control module is used to repeatedly trigger the mating screening module, crossover mutation module, re-evaluation module and screening module to perform operations until the preset number of iterations is reached, and output the optimized target molecular population.

[0081] Example 1 Experimental results comparison and verification: To verify the technical effectiveness of the multi-objective drug molecule generation and optimization system proposed in this application, this embodiment sets up three-objective and four-objective baseline tasks as well as a six-objective extended task. The three-objective and four-objective tasks are compared with published baseline results, and the six-objective task is used for internal comparison of the crossover operator and environment selection strategy.

[0082] It should be noted that the objectives used in the evolution process include model-predicted target activity, QED, SA score, and TPSA and LogP interval preference scores in the six-objective task. Synthetizability, novelty, diversity, and the number of Bemis-Murcko backbones are calculated for external evaluation of the final candidate set and are not directly used as real-time optimization objectives for each generation. In this embodiment, if a generated molecule is determined to be synthesizable based on computational evaluation, it is called a "perfect molecule," and its novelty, diversity, and the number of Bemis-Murcko backbones are further evaluated.

[0083] The baseline results in Tables 1 and 2 use published benchmark data under the same task definition; the methods in this application are calculated according to the same target threshold and evaluation metrics. 5000 molecules were selected from the candidate set that met the corresponding target property requirements for Retrospective analysis. Evaluations were conducted, and novelty, diversity, and skeleton number were calculated on perfect molecules.

[0084] Continuous and discrete types are independent configurations, and their results are listed separately. Tables 3 and 4 show the evaluation of 300 candidate molecules in each group using six-target cross-ablation and environmental selection.

[0085] This embodiment presents the results of a single fixed experiment under various configurations. Data directly referenced from published benchmarks and calculation results from the method of this application are listed in the table by model name.

[0086] This embodiment uses the following four indicators to comprehensively evaluate the final candidate molecule set.

[0087] Number of Bemis-Murcko scaffolds: The Bemis-Murcko scaffolds of the generated molecules are extracted and the number of unique scaffolds is counted to describe the structural coverage of the candidate set at the scaffold level.

[0088] Diversity: Used to measure the structural differences within a set of candidate molecules, calculated based on pairwise Tanimoto distances from Morgan fingerprints.

[0089] in, To generate the total number of drug molecules in the molecular population, and Represents any two distinct drug molecules in the population. drug molecules With drug molecules The similarity between molecules (ranging from 0 to 1, with higher values ​​indicating greater similarity) is calculated using the formula. This formula calculates the similarity between all pairs of molecules, takes the average, and subtracts 1 to obtain a measure of population diversity. The closer the value is to 1, the greater the structural differences between individuals within the molecular population, and the better the diversity.

[0090] Novelty: This measures the structural difference of a candidate molecule relative to a reference set of active molecules. For each candidate molecule, the Morgan fingerprint Tanimoto similarity is calculated between it and its nearest neighbor molecule in the reference set. A candidate molecule is considered novel when the maximum similarity is below 0.4, calculated as follows:

[0091]

[0092] Where N is the number of candidate molecules to be evaluated, G is the number of candidate molecules, R is the set of reference active molecules, sim represents the Tanimoto similarity calculated based on Morgan fingerprints, and I is the indicator function. This index is used to quantify the structural differences between the candidate set and the reference active molecule set.

[0093] Computation of inverse composition solvability (Retro) solvability): Using Retro Perform multi-step inverse synthesis path search. Retro Inverse synthetic programming is modeled as an AND-OR tree search, and a neural network value function guides path expansion. In this embodiment, the maximum number of search iterations is set to 150; if a candidate molecule is decomposed into an available starting building block for the planner within a given search budget, it is denoted as Retro. Solvable; otherwise, unsolvable. (Retro) The solvability rate is the ratio of the number of solvable molecules to the total number of molecules participating in the evaluation.

[0094] Experiment 1: This group of experiments tested the system's performance under three optimization objectives: DRD2, QED, and SA. 5000 molecules were selected from the candidate set that met the objective property requirements for Retrospective analysis. Evaluations were conducted, and novelty, diversity, and skeleton number were calculated on perfect molecules. The experimental results are shown in Table 1. Table 1: Performance Comparison Data of 3D Single-Target Mission Systems

[0095] It should be noted that models with the suffix (F) indicate that a discrete threshold fitness setting is used in the scoring and evaluation phase; PT-MOGA without the suffix (F) uses a continuous scoring setting.

[0096] In a single experiment, PT-MOGA achieved a computational syntheticity of 94.7%, novelty of 61.2%, diversity of 0.815, and a skeleton count of 1378; the corresponding results for PT-MOGA(F) were 88.5%, 86.0%, 0.744, and 767. The discrete setting improved novelty in this task, but simultaneously resulted in a decrease in diversity and skeleton count, indicating that the two scoring methods created different search preferences.

[0097] Experiment 2: This group of experiments targeted four optimization objectives: JNK3, GSK3β, QED, and SA, to test the system's performance in a dual-target, four-objective task. 5000 molecules were selected from the candidate set that met the target property requirements for Retrospective analysis. Evaluations were conducted, and novelty, diversity, and skeleton number were calculated on perfect molecules. The experimental results are shown in Table 2. Table 2: Performance Comparison Data of 4D Dual-Target Mission Systems

[0098] In a single experiment, PT-MOGA achieved a computational synthesability of 75.3%, novelty of 48.0%, diversity of 0.775, and a skeleton count of 1618; the corresponding results for PT-MOGA(F) were 87.3%, 87.0%, 0.722, and 3066. Compared with the published baselines in the table, PT-MOGA(F) achieved higher novelty and skeleton coverage in this task, but its diversity was lower than some baseline methods, indicating that its advantages mainly lie in structural novelty and skeleton expansion after meeting the threshold.

[0099] Example 2 To compare the impact of continuous and discrete fitness settings on search results, this embodiment analyzes the results based on the three-objective task in Table 1 and the four-objective task in Table 2.

[0100] The continuous setting (corresponding to PT-MOGA) preserves the continuous differences in scores for each property; the discrete setting (corresponding to PT-MOGA(F)) converts whether a preset threshold is met to 0 or 1, thereby reducing the selection pressure to continue pursuing extremely high individual scores after the threshold is reached.

[0101] In the three-objective task, the novelty of the continuous and discrete settings was 61.2% and 86.0%, respectively, with 1378 and 767 skeletons, respectively; in the four-objective task, the novelty was 48.0% and 87.0%, respectively, with 1618 and 3066 skeletons, respectively.

[0102] The results show that discrete thresholding improves novelty in both tasks, but its impact on diversity and skeleton number is task-dependent. Therefore, the two scoring methods should be selected based on the actual needs of fine-grained optimization of target properties and exploration of the structure space.

[0103] Example 3 To compare the effects of reaction-sensing crossovers and GB-GA graph structural crossovers on the generation results, this embodiment performed crossover operator ablation under the same six-target NSGA-III environment selection settings. Except for the crossover operator, other major experimental configurations remained consistent, and 300 candidate molecules were selected for evaluation. The experimental results are shown in Table 3.

[0104] Table 3: Comparison of Cross Operator Ablation and Performance in Six-Objective Tasks

[0105] In the experiment, the calculated syntheticity, novelty, diversity, and number of skeletons for GB-GA diagram structural crossovers were 75.0%, 28.0%, 0.7723, and 133, respectively; the corresponding results for molecular converter crossovers were 91.3%, 35.7%, 0.720, and 139, respectively. Molecular converter crossovers achieved higher calculated syntheticity and novelty under this setting, but slightly lower diversity within the set, indicating that reaction-sensing generation, while improving the calculated syntheticity, also resulted in different structural distributions.

[0106] Example 4 To compare the performance of different Pareto environment selection strategies under different objective dimensions, this embodiment employs NSGA-II and NSGA-III in three-objective, four-objective, and six-objective tasks, respectively, and selects 300 molecules from the candidate set of each strategy for evaluation. The experimental results are shown in Table 4.

[0107] Table 4: Performance Comparison Data of NSGA-II and NSGA-III under Different Target Dimensions

[0108] It should be noted that strategy A in the table is for NSGA-II environment selection, and strategy B is for NSGA-III environment selection.

[0109] Experimental results show that different environment selection strategies exhibit task-dependent performance trade-offs. In the three-objective task, NSGA-II has higher novelty and skeleton number than NSGA-III, while their computational synthesizability is similar. In the four-objective task, NSGA-III has higher computational synthesizability and skeleton number, while NSGA-II has higher novelty and diversity. In the six-objective task, NSGA-III has a computational synthesizability of 91.3%, higher than NSGA-II's 75.6%, with novelty of 35.7% and 32.6% respectively, but diversity and skeleton number are similar. These results indicate that NSGA-III's reference direction mechanism is beneficial to improving computational synthesizability under this six-objective setting, but it cannot be inferred that it is superior to NSGA-II in all evaluation metrics and task dimensions.

[0110] Based on the above comparative data, the method of this application has the following technical effects: It can perform a unified process of molecular generation, scoring and multi-objective environment selection in three-objective, four-objective and six-objective tasks, and adopt corresponding diversity preservation strategies according to the objective dimension or preset configuration; In a benchmark evaluation of 5000 molecules, the computational synthesability of the three-target and four-target continuous PT-MOGA converters was 94.7% and 75.3%, respectively. In a six-target comparison of 300 molecules per group, the computational synthesability of the cross-configuration molecular converter using the NSGA-III environment selection was 91.3%. These results indicate that a certain proportion of the generated candidate molecules can be synthesized within a given Retro... The search budget is used to obtain the computational inverse synthesis path; Continuous and discrete scoring, different environment selection strategies, and different crossover operators exhibit different performance trade-offs, which can be configured according to the needs of target property optimization, structural novelty, set diversity, and computational synthesizability.

[0111] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0112] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for multi-target drug molecule generation and optimization, characterized in that, Includes the following steps: S1. Obtain a drug molecule database corresponding to the target point, extract a preset number of drug molecules from the drug molecule database as an initial molecular population, and convert the small molecule compounds in the initial molecular population into a one-dimensional molecular sequence representation to obtain a serialized initial population. S2, perform multi-dimensional fitness evaluation on the initial serialized population to obtain multi-dimensional independent scores for each drug molecule, and aggregate the multi-dimensional independent scores to obtain a scalar reference total score for each drug molecule. The current molecular population is formed by having each drug molecule correspond to its respective scalar reference total score. S3, Based on the scalar reference total score, a set of mating parents is selected from the current molecular population; S4, perform crossover and mutation operations on the parent drug molecules in the mating parent set to generate offspring drug molecules, and merge the parent drug molecules with the offspring drug molecules to form a mixed molecular population; S5, perform a multi-dimensional fitness assessment on the mixed molecular population again to obtain the multi-dimensional independent score corresponding to each drug molecule in the mixed molecular population; S6. Based on the multidimensional independent scores, drug molecules are screened from the mixed molecular population to obtain a new generation of molecular populations. S7. Take the new generation of molecular population obtained in step S6 as the current molecular population for the next iteration, and repeat steps S3 to S6 until the preset number of iterations is reached, and output the optimized target molecular population.

2. The method for multi-target drug molecule generation and optimization according to claim 1, characterized in that, The process of step S2 specifically includes: The original attribute scores of each drug molecule under multiple preset target dimensions are calculated using a corresponding attribute scorer. Based on the preset optimization direction of each evaluation index, the original attribute scores that do not have uniform monotonicity are subjected to monotonicity transformation processing to obtain the uniform monotonicity score of each drug molecule. The monotonicity score of each drug molecule is input into the bimodal fitness conversion mechanism. Based on the optimization task stage or preset instructions, the continuous mode or discrete mode is selected for mapping conversion to form the multidimensional independent score of each drug molecule. The continuous mode retains the continuous floating-point gradient of the monotonic uniform score and performs normalization processing, while the discrete mode sets a corresponding evaluation hard threshold for each preset target dimension and forcibly converts the monotonic uniform score into a discrete binary score of 0 or 1. The multidimensional independent scores corresponding to each drug molecule in each preset target dimension are linearly weighted or summed to obtain the scalar reference total score with a single value.

3. The method for multi-target drug molecule generation and optimization according to claim 1, characterized in that, The process of step S3 specifically includes: Each time, a preset number of drug molecules are randomly selected from the current molecular population as candidates. The scalar reference total score of each candidate is compared. Each time, the candidate with the highest scalar reference total score is selected as a mating parent. This process is repeated until the set of mating parents reaches the preset population size.

4. The method for multi-target drug molecule generation and optimization according to claim 1, characterized in that, Step S4 involves performing a crossover operation on the parent drug molecules in the mating parent set, specifically including: The one-dimensional molecular sequence representation of the paired parent drug molecule is input as the reactant sequence into an attention-based molecular converter model. The molecular converter model predicts the product molecule sequence based on the reactant-to-product mapping relationship learned in the pre-training stage, and the predicted product that passes the structural legality check is used as the candidate progeny drug molecule.

5. The method for multi-target drug molecule generation and optimization according to claim 1, characterized in that, Step S4 involves performing mutation operations on the parent drug molecules in the mating parent set, specifically including: According to the preset mutation probability, the molecular structure of drug molecules is applied to a preset chemical reaction template composed of reaction rules defined by SMARTS, so as to achieve local fine-tuning of the molecular structure.

6. The method for multi-target drug molecule generation and optimization according to claim 1, characterized in that, The process of step S6 specifically includes: Based on the multidimensional independent scores, non-dominated sorting is performed on the mixed molecular population to construct multiple Pareto levels; The number of preset target dimensions in step S2 is used as the number of optimization targets and compared with a preset dimension threshold. For drug molecules within the same Pareto level, if the number of the preset target dimensions is less than or equal to the dimension threshold, a diversity maintenance strategy based on a combination of crowding distance and random exploration is executed; if the number of the preset target dimensions is greater than the dimension threshold, a niche diversity assessment strategy based on reference point association is executed; and the next generation of molecular populations is screened based on the executed strategy.

7. The method for multi-target drug molecule generation and optimization according to claim 6, characterized in that, The specific process of implementing the diversity maintenance strategy based on a fusion of crowding distance and random exploration is as follows: Combining the non-dominated ranking order and crowding distance of drug molecules, elite drug molecules at the forefront are preferentially screened according to the first preset retention quantity; Random sampling is performed from the remaining non-elite drug molecules in the mixed molecular population, according to a second preset retention quantity; The selected elite drug molecules are combined with randomly sampled drug molecules to form the new generation of molecular population. The specific process for implementing the reference point-based microhabitat diversity assessment strategy is as follows: The priority order of drug molecules is determined based on their non-dominated ordering. Calculate the correlation distance between each drug molecule and a pre-defined reference point in the target space; Based on the correlation distance and niche count results, drug molecules within the same Pareto level are screened; A predetermined number of drug molecules are extracted from the mixed molecular population to form the new generation molecular population.

8. The method for multi-target drug molecule generation and optimization according to claim 2, characterized in that, The attribute scorer includes at least one of the following: a target affinity predictor based on machine learning, a heuristic calculator based on physicochemical rules of thumb, and a statistical scorer based on historical reaction fragments. The monotonicity transformation process includes: applying a linear extreme value mapping and inversion formula to the original score for the one-sided extreme value constraint attribute to convert the original score into a normalized score where the larger the value, the better the performance; and introducing a bell-shaped nonlinear normalization formula based on the Gaussian distribution function for the two-sided interval constraint attribute. When the original score of the drug molecule hits the preset optimal target bias, the score is full; if it deviates from the optimal target bias, the score smoothly decays towards zero.

9. A multi-target drug molecule generation and optimization system, characterized in that, include: The database acquisition module is used to acquire a database of drug molecules corresponding to the target. An initialization module is used to extract a preset number of drug molecules from the drug molecule database as an initial molecular population, and to convert the small molecule compounds in the initial molecular population into a one-dimensional molecular sequence representation to obtain a serialized initial population. The evaluation aggregation module is used to perform multi-dimensional fitness evaluation on the initial serialized population, obtain multi-dimensional independent scores for each drug molecule, and aggregate the multi-dimensional independent scores to obtain a scalar reference total score for each drug molecule. The current molecular population is constructed in such a way that each drug molecule corresponds to its respective scalar reference total score. The mating screening module is used to screen out a set of mating parents from the current molecular population based on the scalar reference total score; The crossover and mutation module is used to perform crossover and mutation operations on the parent drug molecules in the mating parent set to generate offspring drug molecules, and to merge the parent drug molecules with the offspring drug molecules to form a mixed molecular population; The re-evaluation module is used to re-evaluate the multi-dimensional fitness of the mixed molecular population to obtain the multi-dimensional independent score corresponding to each drug molecule in the mixed molecular population. A screening module is used to screen drug molecules from the mixed molecular population based on the multidimensional independent scores to obtain a new generation of molecular populations. The loop control module is used to repeatedly trigger the mating screening module, the crossover mutation module, the re-evaluation module, and the screening module to perform operations until a preset number of iterations is reached, and output the optimized target molecular population.

10. A computer device, characterized in that, include: Processor and computer-readable storage media; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements a multi-target drug molecule generation and optimization method as described in any one of claims 1 to 8.

Images