Drug design and screening method and system based on group intelligent weak meat strong food algorithm

Abstract

The invention discloses a drug design optimization and precision medical analysis method and system based on swarm intelligence. According to the method, behaviors of weak meat and strong food groups are simulated, candidate drug molecules or patient pathological models are regarded as agents, and core data of the agents are protected by hiding virtual models; risks such as toxicity and off-target effect are actively identified and avoided through the security defense model; the intelligent agent is guided to escape from the local optimal solution and the invalid research and development path through the escape model; and finally performing fusion and multiple verification on the elite agent through the survival model, and outputting high-quality and low-risk candidate drug molecules or personalized treatment schemes. According to the method, a safe and anti-interference mechanism is embedded in an optimization process, the efficiency and success rate of new drug research and development are remarkably improved, and a powerful analysis tool is provided for precision medical treatment.

Description

Technical Field

[0001] This invention relates to the field of computational medicinal chemistry and artificial intelligence, and in particular to a swarm intelligence algorithm that simulates the law of survival of the fittest in nature, and its application in virtual drug screening, lead compound optimization and de novo drug design. Background Technology

[0002] Drug discovery is a time-consuming, costly, and highly unsuccessful process. Traditional drug discovery relies heavily on extensive experimental screening and experience-based chemical modifications. With the development of computer science, computer-aided drug design has become an indispensable tool, with structure-based virtual screening and quantitative structure-activity relationship models being widely used.

[0003] However, existing computational drug design methods face many challenges:

[0004] 1. The vastness of the chemical space: The number of possible drug molecules is estimated to be on the order of 10^60. Existing calculation methods are difficult to explore this huge space efficiently and comprehensively, and are prone to getting stuck in local optima and missing out on high-quality candidate molecules with novel structures.

[0005] 2. The complexity of multi-objective optimization: An ideal drug candidate needs to achieve a balance across multiple dimensions, including potency (binding affinity to the target), selectivity, pharmacokinetics (absorption, distribution, metabolism, excretion), toxicity, and synthetic feasibility. Simultaneously optimizing these often conflicting objectives is extremely difficult.

[0006] 3. The "vulnerability" of the computational process: During optimization, the algorithm may generate molecules that theoretically score highly but are actually chemically unstable, difficult to synthesize, or perform poorly in complex organisms. These molecules can be considered "invaders" or "traps" in the computational process, as they can mislead the search direction and waste computational resources.

[0007] While existing evolutionary algorithms (such as genetic algorithms) have applications in drug design, they typically lack mechanisms to dynamically address the aforementioned "traps." Like a population without defenses, these algorithms are easily swayed by undesirable individuals, leading to the failure of the entire optimization process.

[0008] Therefore, there is an urgent need in this field for a novel intelligent algorithm with inherent robustness and dynamic defense capabilities. This algorithm can not only efficiently search a vast chemical space, but also actively identify and avoid chemical and biological "traps" during the search process, thereby discovering candidate molecules with truly high drug potential in a more intelligent and reliable manner. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for applying a survival-of-the-fittest swarm intelligence algorithm to drug design. By simulating the core principles of "hide-defense-escape-survival" in an ecosystem, virtual drug candidate molecules possess self-protection and evolutionary capabilities during the optimization process, thereby more efficiently and reliably screening lead compounds with optimal overall properties.

[0010] Technical solution

[0011] To achieve the above objectives, this invention proposes a drug design and screening method based on a swarm intelligence survival-of-the-fittest algorithm. Its core principle is to treat each virtual candidate molecule as an intelligent agent striving for survival in a complex "pharmaceutical environment" (a multidimensional property space).

[0012] The algorithm consists of four core behavioral models and is deeply integrated with the professional requirements of drug design:

[0013] 1. Hidden and Virtual Models: At the start of the algorithm, the true chemical identity of each candidate molecular agent (such as its precise 3D structure and SMILES sequence) is converted into a hashed or encrypted "virtual ID card." This has two advantages: first, it protects intellectual property in distributed computing, preventing the leakage of molecular structures; second, at the algorithm level, it avoids the search process from becoming prematurely "obsessed" with certain locally optimal structural features, prompting the algorithm to explore a more fundamental "property space" rather than simply a structural space.

[0014] 2. Safety Defense Model: When candidate molecules explore the property space, they encounter various "aggressions." For example, a molecule might twist into an energetically unstable conformation in pursuit of an extremely high docking score (conformational aggression); or a molecular fragment might cause severe predicted toxicity (toxicity aggression). The safety defense model monitors these threats in real time through built-in "molecular mechanics functions" and "toxicity prediction functions." Once detected, it triggers defense mechanisms, such as guiding the molecule to make minor conformational adjustments to restore stability, or evading toxicity by replacing specific fragments.

[0015] 3. Escape Model: When the defense model is insufficient to solve the problem (for example, no matter how a molecule is adjusted, its core skeleton is difficult to avoid binding to a certain undesirable secondary target), the escape model is activated. This is no longer fine-tuning, but a more significant "strategic shift." Based on functions such as free energy calculation and functional group substitution probability, this model guides candidate molecules to perform "skeleton jumps" or "major pharmacophore substitutions," thereby fundamentally escaping the current "danger zone" and jumping to a completely new and safer subregion in the chemical space for exploration.

[0016] 4. Survival Model: This is the convergence and elitism stage of the algorithm. After multiple rounds of brutal "property screening," the surviving candidates are all "strong" individuals with balanced and excellent performance across various indicators.

[0017] The system employs "elite management" of these molecules.

[0018] Alliances and mergers: Allowing these excellent molecules to exchange properties, such as combining the highly active pharmacophore of one molecule with the excellent pharmacokinetic skeleton of another molecule to form a new and more powerful "alliance molecule".

[0019] Multiple validations: These final candidates undergo the most rigorous cross-validation, using multiple different scoring functions, more accurate molecular dynamics simulations, and comprehensive predictions based on deep learning-based ADMET, to grant them a "drug-likeness pass" and greatly improve their success rate in entering the physical experiment stage.

[0020] Beneficial effects

[0021] Compared with the prior art, the present invention has the following significant advantages:

[0022] 1. Higher search efficiency and novelty: By escaping model-driven skeleton transitions, the algorithm can break through local optima and explore candidate molecules with more novel structures and greater patent space, which is difficult to achieve with traditional methods.

[0023] 2. Inherent robustness and reliability: The algorithm has an "error prevention" mechanism that can actively avoid molecular traps with unreasonable chemistry and poor properties, making the entire optimization process more robust and the output results more reliable.

[0024] 3. Excellent multi-objective optimization capability: It integrates multiple objectives such as efficacy, pharmacokinetics, and toxicity into the survival pressure of the algorithm, naturally forcing candidate molecules to evolve towards a comprehensive balance rather than the extreme of a single indicator.

[0025] 4. End-to-end intelligence: The experience of medicinal chemists (such as avoiding certain toxic structures) is transformed into defense and evasion rules built into the algorithm, realizing the automation and intelligent integration from "chemical space exploration" to "drugability assessment", which significantly accelerates the process of preclinical research. Attached Figure Description

[0026] Figure 1 The following is a flowchart illustrating the overall process of lead compound optimization using the method described in this embodiment of the invention.

[0027] Figure 2 This invention illustrates how candidate molecular agents are optimized through defense and evasion behaviors in a two-dimensional "activity-toxicity" property space. Figure 3 3D rendering.

[0028] Figure 3 : A schematic diagram of the optimization of candidate molecular intelligence in the "activity-toxicity" two-dimensional property space through defense and evasion behaviors in this invention.

[0029] Figure 4 : A schematic diagram of property merging using the "alliance algorithm" in the survival mode described in this invention.

[0030] Figure 5 The flowchart of the property merging process of the "alliance algorithm" in the survival mode described in this invention. Detailed Implementation

[0031] The application of the present invention in the scenario of "optimizing the hepatotoxicity of known lead compounds" will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] Background: A lead compound A exhibits high activity against its target, but computer analysis predicts potential hepatotoxicity (e.g., inhibition of CYP450 enzymes). Our goal is to find novel compounds that significantly reduce hepatotoxicity while maintaining activity.

[0033] System General Configuration and Data Preprocessing

[0034] Before describing specific embodiments, the general technical infrastructure required to implement this invention is first defined:

[0035] 1. Computing hardware platform:

[0036] CPU: At least 16 cores, with a clock speed of 2.5GHz or higher, used for molecular dynamics simulations and parallel computing.

[0037] GPU: NVIDIA Tesla V100 or equivalent or higher computing power, used to accelerate deep learning model inference (such as toxicity prediction, property prediction) and some search algorithms.

[0038] Memory: at least 128 GB.

[0039] Storage: High-speed SSDs with a capacity of at least 2TB are used to store large compound libraries and intermediate results.

[0040] 2. Core Software and Dependency Libraries:

[0041] Cheminology Fundamentals: RDKit (2022.09 or later) for reading, processing, fingerprinting, and calculating simple properties of molecules.

[0042] Molecular docking: AutoDock Vina (1.2.3) or QuickVina 2 for rapid binding affinity assessment; commercial software such as Schrödinger Suite (Glide module) can be used for high-precision verification.

[0043] Molecular dynamics simulations: GROMACS (2022 or later) or OpenMM, for conformational stability assessment.

[0044] Machine learning frameworks: PyTorch (1.12 or later) or TensorFlow (2.10) for building and running custom defense, evasion and prediction models.

[0045] ADMET Prediction Platform: Locally deployed ADMETlab 2.0 API or DeepPurpose toolkit.

[0046] 3. Basic database preparation:

[0047] Safe / Low Toxicity Fragment Library: Approximately 500 common pharmacophore fragments with LogP < 3 and no warning structures (such as nitro or Michael receptor) are selected from databases such as ChEMBL and PubChem and stored as SMARTS mode or SMILES list.

[0048] Bioisosteric substitution rule library: It organizes recognized bioisosteric pairs, such as "benzene ring ↔ thiophene", "carboxylic acid ↔ tetrazolium", "amide ↔ urea", etc., to form a searchable dictionary structure.

[0049] Example 1: Optimization of hepatotoxicity of a known lead compound (corresponding to claim 8)

[0050] This embodiment details how the present invention can be used to reduce the predicted hepatotoxicity of lead compound A while maintaining its target activity.

[0051] Step S101: Initialize the population (concretization)

[0052] 1. Input the specification for lead compound A

[0053] SMILES: CN1C(=O)C2=CC=CC=C2C(=O)N1C3=CC=C(C=C3)O`.

[0054] 2. Generate analogues:

[0055] Parse using `Chem.MolFromSmiles` from RDKit.

[0056] Derivation strategy: The following rules are used to generate the initial population (N=300 individuals in total) in batches via script:

[0057] Ring modification: Expand (5→6), shrink (6→5), or open the non-aromatic ring, with a probability of 10% for each.

[0058] Substituent substitution: Hydrogen atoms on the benzene ring are randomly replaced with -F, -Cl, -CH3, -OCH3, -OH, -CF3, with a substitution probability of 15% at each site.

[0059] Linker variation: Replace the amide linker (-CONH-) with sulfonamide (-SO2NH-), urea (-NHCONH-), or trans-amide, with a probability of 10%.

[0060] 3. Assign hidden virtual identities:

[0061] For each generated molecule, calculate its Morgan fingerprint (radius=2, length=2048 bits).

[0062] The fingerprint byte stream is hashed using the SHA-256 algorithm from Python's hashlib library, generating a 64-bit hexadecimal string as the agent's unique hidden ID (e.g., e3b0c44298fc1c14...). This ID represents the molecule throughout the algorithm iterations, recording its trajectory and properties without revealing its true structure.

[0063] 4. Conformation generation and optimization: For each molecule, up to 5 initial conformations are generated using RDKit's EmbedMultipleConfs, and minimized using the MMFF94 force field. The conformation with the lowest energy is selected as its representative 3D structure and saved in PDBQT format for docking.

[0064] Step S102: Conduct a global exploration (concretization) in the pharmaceutical property space.

[0065] 1. Defining a multidimensional pharmaceutical property space: This embodiment focuses on two core dimensions and one auxiliary dimension.

[0066] Dimension 1 (Activity Axis): Binding free energy (ΔG, kcal / mol) with target protein X. Calculated using AutoDockVina within a uniform grid box (centered at the known ligand binding site, size 20 Å × 20 Å × 20 Å). The scoring function is the default Vina function.

[0067] Dimension 2 (Toxicity Axis): Predicted probability values ​​(0-1) for hepatotoxicity (CYP3A4 inhibition). Predictions were made using a pre-trained graph neural network (GNN) model. This model was trained on a CYP-related task on the `Tox21` dataset with AUC > 0.85.

[0068] Dimension 3 (Feasibility Axis): Composite Accessibility Score (SA Score, 1-10, lower is better). Calculated using the `CalcSAScore` function built into RDKit.

[0069] 2. Exploring Algorithms – Improved Multi-Objective Particle Swarm Optimization (MOPSO):

[0070] Particle encoding: The position of each particle (corresponding to a molecular agent) is a 3-dimensional vector [norm(ΔG),norm(Toxicity),norm(SA)], wherenorm() is a function that linearly normalizes the original value to the interval [0,1].

[0071] Velocity and position updates: follow the standard PSO formula. Key parameters are set as follows: population size M = 300, inertia weight w decreases linearly from 0.9 to 0.4, and acceleration constants c1 = c2 = 1.496.

[0072] Exploration Strategy: Introduce the concept of simulated annealing, and accept an "exploratory move" with a 10% probability even if all objectives deteriorate, to avoid getting stuck in a local Pareto front.

[0073] Iteration control: Maximum number of iterations T_max = 50. Convergence criterion: If the hypervolume index of the entire population changes by less than 1% in 10 consecutive iterations, the exploration is terminated early.

[0074] Step S103: Security Defense Model (Concretization)

[0075] 1. Aggression Detection Trigger: In each iteration, after evaluating each particle, check the following condition:

[0076] Condition 1 (Conformational Invasion): If the molecule's "movement" (i.e. structural modification) in this iteration to improve the docking score results in the breaking of key hydrogen bonds or severe disruption of aromatic ring coplanarity (as assessed by RDKit's CalcTorsionalStrain, the torsion angle energy increases by >5 kcal / mol), then it is determined to have suffered conformational invasion.

[0077] Condition 2 (Toxic invasion): If the predicted probability value of hepatotoxicity exceeds the threshold Th_tox=0.7, then it is determined that toxic invasion has occurred.

[0078] 2. Execution of defensive actions:

[0079] For conformational invasion: Initiate the "Conformation Repair" subroutine. Using Monte Carlo simulation combined with the MMFF94 force field, randomly rotate (within ±30°) the bonds that have undergone unfavorable torsion, sample 10 times, and select the conformation with the lowest internal energy that does not significantly reduce the docking score (ΔΔG < 1.5 kcal / mol) for replacement.

[0080] To address toxicity invasion: Initiate the "Toxicity Fragment Replacement" subroutine. Systematically analyze which substructures in the molecule are highly correlated with the toxicity prediction model (using the Grad-CAM gradient interpretation method of GNN). After identifying the suspect fragment, randomly select three fragments with similar shapes and polarities from a pre-set safe fragment library for equal replacement, reassess the toxicity, and select the scheme with the greatest reduction in toxicity and the least loss of activity.

[0081] Step S104: Escape Model (Concretization)

[0082] 1. Activation condition: When the same molecular agent triggers toxicity defense for 3 consecutive iterations and the toxicity value is still higher than `Th_tox` after defense, the region is determined to be a "toxicity dead zone" and the escape model is forcibly activated.

[0083] 2. Escape Policy Computation – Reinforcement Learning-Based Skeleton Transfer:

[0084] State: The Morgan fingerprint (ECFP4) of the current molecule.

[0085] Action: A predefined set of chemical transformations, such as: {'benzene ring -> pyridine ring', 'amide -> sulfonamide', 'removal of the methyl group', 'addition of a piperidine ring'}. This example predefines 20 such actions.

[0086] Reward: R = -ΔToxicity + 0.5 * (-ΔΔG) - 0.1 * ΔSA. That is, decreased toxicity and maintained activity are positive rewards, while increased synthesis difficulty is a negative reward.

[0087] Training and Decision-Making: A lightweight deep Q-network (DQN) is used. It is pre-trained on a small molecular dataset before the algorithm runs. When escape is activated, the network takes the current molecular state as input, outputs the Q-values ​​of all possible actions, and selects the action with the highest Q-value to execute.

[0088] 3. Example of Escape from Execution: Suppose the network determines the optimal action is "benzene ring -> pyridine ring". The system replaces the identified benzene ring SMARTS pattern c1ccccc1 in the molecule with pyridine n1ccccc1, generating a new molecule, and then performs 3D conformation generation, optimization, and property evaluation again. This molecule has thus completed a "strategic position shift", and its property vector will undergo a sudden change.

[0089] Step S105: Survival Mode and Coalition Algorithm (Specification)

[0090] 1. Strongest selection: After the exploration phase, select all individuals on the non-dominated solution (Pareto front) from the final population and define them as "strongest agents" (assuming there are K=15).

[0091] 2. Alliance generation (nature-based annexation):

[0092] Pairing: Randomly pair the strongest agents together to generate C(K,2) pairs.

[0093] Cutting and splicing: For each pair of molecules (M1, M2):

[0094] (1) Use the BRICS algorithm (implemented by RDKit) to identify the cuttable bond between the two, usually choosing to cut near the ring system or at the flexible connector.

[0095] (2) Join a fragment of M1 with a complementary fragment of M2. For example, retain the core framework and active pharmacophore of M1 and join the polar side chain of M2 to improve solubility.

[0096] (3) Use reaction templates (such as amide bond formation) to create virtual linking fragments to ensure the generation of chemically sound molecules.

[0097] In this embodiment, L=2 alliance molecules are generated for each pair of strong players, resulting in a total of approximately 15*14 / 2*2=210 new candidate molecules.

[0098] 3. Multiple verification system:

[0099] First step: Cross-docking verification: Two different docking programs, AutoDock Vina and LeDock, were used to dock 210 consortium molecules. Only molecules with binding free energies better than -9.0 kcal / mol in both programs were retained (approximately 50 molecules remaining).

[0100] Second step: Molecular dynamics stability verification: For the above 50 molecules, complexes were constructed with target proteins, and short molecular dynamics simulations (GROMACS, CHARMM36 force field) were performed for 10 ns. The protein-ligand centroid distance and ligand RMSD within 2 ns after the simulation were calculated. Complexes with stable distances (fluctuations <2 Å) and RMSD <3 Å were retained (approximately 15 remaining).

[0101] The third step: Comprehensive ADMET virtual screening: Fifteen molecules were submitted to the ADMETlab 2.0 platform for prediction of over 30 properties. Hard filtering conditions were set: hepatotoxicity probability <0.3, low hERG risk, solubility (LogS) >-5, and no inhibition of cytochrome P450 2D6. Molecules that passed all conditions were included in the final list.

[0102] Step S106: Output the result

[0103] 1. Sorting and Report Generation: The final candidate molecules that pass the triple validation (e.g., 3) are sorted according to their comprehensive score Z = 0.6*(-ΔG) + 0.3*(1-Toxicity) + 0.1*(10-SA_Score).

[0104] 2. Output content:

[0105] File 1 (Recommended Compounds): Contains the specification SMILES, hidden ID (traceable), predicted ΔG, toxicity value, SA score, and detailed data tables (CSV format) of the top 3 molecules in each round of validation.

[0106] File 2 (3D structure): Optimized 3D conformation of the ranked molecule and its docking pattern with the target (PDB format).

[0107] File 3 (Synthesis Suggestions): 1-3 possible retrosynthetic routes generated for the molecule ranked 1st, based on the AiZynthFinder or ASKCOS tools.

[0108] 3. Conclusion of Example 1: Through the above process, this invention starts from 300 analogues and finally outputs 2-3 optimized candidate compounds that, while maintaining nanomolar activity, reduce the predicted hepatotoxicity probability from the initial >0.7 to <0.3, and includes their detailed property profiles and preliminary synthetic feasibility analysis.

[0109] Example 2: High-throughput virtual screening for specific targets (corresponding to claim 6)

[0110] (This section will use the same level of detail as Example 1, but with a different focus.)

[0111] Step S101 Differentiation: The initial population consists of 100,000 molecules randomly selected from the ZINC15 "drug-like" subset. The method for generating hidden identities is the same.

[0112] Step S102 Differentiation: The pharmaceutical property space is simplified to two dimensions: 1) predictive affinity with disease targets (using the ultrafast docking tool QuickVina 2); 2) drug-likeness (QED score). The exploration algorithm employs a tree-based Pareto Estimator (TPE) for Bayesian optimization to efficiently handle large-scale initial populations.

[0113] Step S103 / S104 Differentiation: The defense and evasion model is mainly aimed at "drug-like aggression", such as when a molecule becomes too complex in pursuit of high activity (SA Score deterioration) or violates the "five rules", triggering defense (simplification of structure) or evasion (replacement of complex fragments).

[0114] Step S105 Differentiation: The consortium algorithm is performed on the Top 500 active molecules selected, focusing on pharmacophore hybridization to enhance activity.

[0115] Output: The final output is a list of 10-20 highly active lead compounds with novel skeletons.

[0116] Example 3: Fragment-based de novo drug design (corresponding to claim 7)

[0117] (This section will use the same level of detail as Example 1, but with a different focus.)

[0118] Step S101 Differentiation: The initial population consists of 500 small molecule fragments (molecular weight <250 Da), which are obtained from commercial fragment libraries or by virtual cutting of known drugs.

[0119] Step S102 Differentiation: Global exploration takes place in a "reaction-based synthetic space." Each agent's "position" is defined not only by properties but also by a list of possible chemical reactions. The exploration algorithm drives virtual chemical reactions (such as amide synthesis and Suzuki coupling) between fragments, progressively growing molecules.

[0120] Steps S103 / S104 Differentiation: The defensive model focuses on defending against "synthetic infeasibility aggression," that is, preventing the algorithm from generating intermediates that are difficult to synthesize. The evasion model, on the other hand, allows molecules to "backtrack" a few steps and choose different reaction pathways when the growth path is unfavorable.

[0121] Step S105 Differentiation: The survival mode emphasizes skeleton novelty checks (by querying public databases) to ensure that the final molecule has patent space.

[0122] Output: The final output consists of 5-10 novel, synthetic molecular entity designs that meet the property prediction requirements.

Claims

1. A drug design and screening method based on a swarm intelligence survival-of-the-fittest algorithm, characterized in that, Includes the following steps: Step S101: Initialize a population consisting of multiple virtual drug candidate molecules. Each candidate molecule is treated as an intelligent agent and given a mathematically transformed hidden virtual identity. This identity is used to conceal its real chemical descriptor and structural information during the calculation process. Step S102: The candidate molecule agent performs a global exploration in a predefined multidimensional pharmaceutical property space, which includes at least molecular docking affinity, pharmacokinetic properties, toxicity parameters, and synthetic feasibility dimensions. Step S103: During the exploration process, each candidate molecule agent activates a security defense model to identify and defend against "molecular invasion" from other candidate molecules or the external simulated environment through a first preset function. The molecular invasion includes interference with its chemical stability and disruption of its binding conformation with the target. Step S104: When the security defense model determines that it cannot effectively resist the current invasion, and the invasion will cause unacceptable property degradation of the candidate molecular agent, the escape model is activated. The second preset function is used to calculate and execute a strategic position transfer in the pharmaceutical property space in order to escape the invasion and find a new stable region. Step S105: After multiple rounds of exploration, defense and evasion, the surviving candidate molecular agents enter the survival mode. The survival mode establishes a multi-verification system for the strongest agents with the best properties, and integrates and strengthens the superior properties through the alliance algorithm. Step S106: Output the molecular structure and pharmaceutical properties of the candidate molecular agents that ultimately survive in the survival mode, as lead compounds with high drugability.

2. The method according to claim 1, characterized in that, The hidden virtual identity in step S101 is achieved by generating an anonymous identifier through feature hashing or homomorphic encryption of the molecule's SMILES string or three-dimensional structural coordinates.

3. The method according to claim 1, characterized in that, The security defense model in step S103 specifically includes: The conformational stability of candidate molecules under simulated physiological conditions is evaluated using molecular dynamics simulation functions. When the stability is below a threshold, conformational adjustment is triggered to resist "unfolding invasion". Virtual target decoys are activated to guide "aggressive" molecular fragments with non-specific binding tendencies to unrelated virtual targets, thereby protecting the optimal binding mode of candidate molecules to real targets.

4. The method according to claim 1, characterized in that, The escape model in step S104 adopts an escape strategy based on free energy change calculation. The second preset function is a multivariate function based on the binding free energy of the current candidate molecule, the free energy change of aggressive interference, and the physicochemical properties of the environment, used to calculate its conformational adjustment, side chain rotation, or skeletal transition escape path.

5. The method according to claim 1, characterized in that, The survival mode in step S105 further includes: Fragment exchange or pharmacophore splicing is performed between strong candidate molecule agents to achieve "property annealing" at the algorithm level, so as to generate a new generation of candidate molecules with multiple excellent properties; The multi-validation system includes: cross-validation using multiple independent scoring functions, and a second round of virtual screening using an AI-based ADMET prediction model.

6. The method according to claim 1, characterized in that, The method is used for high-throughput virtual screening of specific disease targets, wherein the objective function in the pharmaceutical property space is the predicted binding affinity of candidate molecules to the disease target.

7. The method according to claim 1, characterized in that, The method is used for de novo drug design, wherein the global exploration in step S102 is carried out in a fragment-based chemical space or a reaction-based synthetic space to generate novel molecular entities with patent novelty.

8. The method according to claim 1, characterized in that, The method is used to optimize the property defects of a known lead compound, wherein the initial population consists of structural analogs or derivatives of the lead compound, and the objective function is the degree of optimization of a specific toxicity parameter or pharmacokinetic parameter while maintaining activity.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs system identification, reading, and storage according to the method described in any one of claims 1 to 8.

Images