A Method for Identifying Effective Small Molecules in Traditional Chinese Medicine Prescriptions by Integrating EGA and PPO Algorithms

Abstract

This invention discloses a method for identifying effective small molecules in traditional Chinese medicine (TCM) prescriptions by integrating EGA and PPO algorithms. It relates to the field of small molecule identification technology for TCM decoctions, and includes constructing a molecular seed library containing known components from TPACD and structurally similar phytochemicals. EGA simulates the selection, transformation, and recombination mechanisms during the TCM decoction process, while PPO leverages the sequential decision-making optimization capabilities of reinforcement learning. Three parallel modules—Basic Multi-Objective Genetic Algorithm (BMGA), Enhanced Multi-Objective Genetic Algorithm (EMGA), and Intelligent Multi-Objective Genetic Algorithm (IMGA)—are built. Through molecular evolution, targeted screening, iterative refinement, and closed-loop feedback, anti-CRC leader molecules with efficacy, stability, and drug-like properties are screened. The molecules generated by this invention significantly outperform the original TPACD components in terms of drug-likeness, structural stability, and anti-CRC activity. The IMGA module shows the best overall performance, providing an efficient and interpretable new strategy for mining active ingredients in TCM compound prescriptions.

Description

Technical Field

[0001] This invention relates to the field of small molecule discovery algorithms for traditional Chinese medicine decoctions, specifically to a method for identifying effective small molecules in traditional Chinese medicine prescriptions that integrates EGA and PPO algorithms. Background Technology

[0002] Colorectal cancer (CRC) is one of the most prevalent and deadliest malignant tumors worldwide, placing a heavy burden on public health. Despite advancements in surgical treatment, chemotherapy, and targeted therapy, tumor heterogeneity, drug resistance, and treatment-related systemic toxicity continue to significantly impact clinical prognosis. Traditional Chinese medicine (TCM) compound formulas, as an important form of TCM treatment, demonstrate unique potential in adjuvant therapy for CRC due to their advantages of multiple components, multiple targets, and low toxicity. Among them, Tiaopi Anchang Decoction (TPACD) has been clinically validated for its significant anti-CRC effect. Combined UHPLC-Q-TOF-MS / MS and network pharmacology analysis revealed that it may target matrix metalloproteinases (such as MMP3) as a potential therapeutic target. However, due to its extremely complex chemical composition, the exact active ingredients and molecular mechanism of action remain unclear.

[0003] Traditional methods for screening active ingredients in traditional Chinese medicine (TCM) mainly rely on activity-tracking separation strategies. These methods suffer from low throughput, strong structure bias, and time-consuming processes, making them unsuitable for the systematic analysis of multi-component TCM formulas. In recent years, the rapid development of computational chemistry and artificial intelligence has provided new technical pathways for drug molecule design. Generative models and de novo molecular design algorithms can now create compounds with novel structures and controllable pharmacological properties. However, directly applying these methods to the chemical space of TCM still faces many challenges: ensuring the chemical efficacy of the generated molecules, preserving the unique skeletal characteristics of TCM components, balancing molecular novelty and drug-likeness, and capturing the core logic of the synergistic effects of multiple components in TCM formulas.

[0004] Therefore, there is an urgent need for a method for identifying effective small molecules in traditional Chinese medicine prescriptions that integrates EGA and PPO algorithms. This method can simulate the dynamic process of "processing-compatibility-refining" in traditional Chinese medicine compound prescriptions and achieve accurate screening of active molecules through efficient computational optimization, thus providing a new solution for the discovery and development of active ingredients in traditional Chinese medicine compound prescriptions. Summary of the Invention

[0005] This invention provides a method for identifying effective small molecules in traditional Chinese medicine prescriptions that integrates EGA and PPO algorithms. It aims to solve the problems of low screening efficiency, blind optimization of structure, and difficulty in balancing activity and drug-likeness in existing technologies. The method screens out candidate molecules for colorectal cancer that have high activity, good drug-likeness, and structural stability, providing new lead compounds for CRC treatment. At the same time, it provides a scalable and interpretable technical framework for the mining of active ingredients in complex traditional Chinese medicine prescriptions.

[0006] The method for identifying effective small molecules in traditional Chinese medicine prescriptions by integrating EGA and PPO algorithms provided by this invention specifically includes the following steps:

[0007] S1. Obtain the known components of the traditional Chinese medicine Tiaopi Anchang Decoction (TPACD) and phytochemicals with structures similar to the known components, forming an original compound set containing more than 1,000 compounds.

[0008] S2. Standardize and protonate the molecules in the original compound set, and eliminate simple structures by complexity screening to construct a molecular seed library;

[0009] S3. Extract molecules from the molecular seed bank and initialize them to form an initial population;

[0010] S4. Input the initial population into a computational system that integrates the Evolutionary Genetic Algorithm (EGA) and the PPO reinforcement learning algorithm. Through continuous molecular evolution, targeted screening, iterative refinement, and closed-loop feedback, a set of candidate molecules that meet the requirements is obtained.

[0011] The computing system includes the basic multi-objective genetic algorithm BMGA, the enhanced multi-objective genetic algorithm EMGA, and the intelligent multi-objective genetic algorithm IMGA.

[0012] Preferably, the specific content of S4 includes:

[0013] S401. Start the BMGA module and evolve the initial population through simplified mutation and crossover operators. After selection and convergence, obtain the structurally diverse BMGA output molecular backbone.

[0014] S402. Input the BMGA output molecular skeleton into the EMGA module, introduce affinity prediction and drug-likeness scoring as guidance, and perform selection and convergence again to guide the evolutionary process to a chemically relevant chemical spatial region to form an EMGA high-potential molecule set.

[0015] S403. Select molecules with good performance from the EMGA high-potential molecule set as seed molecules, input them into the IMGA module, and perform iterative refinement through PPO reinforcement learning to obtain the IMGA optimized molecule set.

[0016] S404. Feed the optimized IMGA molecular set back to the BMGA module to enrich the EGA population. At the same time, dynamically feed the screening results of EMGA back to IMGA, adjust the reward weight of reinforcement learning, and repeat S401 to S404 until the iterative refinement stopping condition is met.

[0017] In each generation of evolution, candidate molecules are filtered for chemical stability, novelty, and drug-likeness, and ineffective SMILES, PAINS positive, or highly toxic structures are penalized.

[0018] Preferably, the evolutionary process of the molecular population follows a multi-objective optimization principle, expressed as:

[0019] ;

[0020] in, For the t-th generation molecule to be evaluated Overall score QED This is a quantitative estimate of the drug-like properties. SA Represents the accessibility of synthesis. To predict affinity, For the structural novelty based on the Tanimoto distance, Here are the weighting coefficients for each indicator, and the weights... After normalization, it satisfies .

[0021] Preferably, the EGA algorithm employs a dynamic tournament selection strategy to "select" the heat control characteristics of the simulated decoction, adjusting the selection pressure across different generations. ρ , ρ It grows exponentially, gradually prioritizing individuals with higher fitness; early generations emphasize diversity, for low... ρ Later generations focused on convergence, for high... ρ ;

[0022] The formula for calculating the probability of selection is:

[0023] ;

[0024] in, No. t molecule The probability of being selected to participate in subsequent evolution, where p is the selection pressure parameter. Let be the fitness value of the t-th molecule. It is the sum of the fitness values ​​of all molecules in the current population raised to the power of p.

[0025] Preferably, the characteristic is that "crossing" corresponds to the compatibility stage of herbal processing, generating new functional relationships through the combination of components, and the expression for the crossing operation is:

[0026] ;

[0027] in, This is a graph of offspring molecules generated through crossover operations. This is a crossover operation function. , Molecular diagrams of the two parent molecules involved in the crossover. For fragment assembly functions, and From the father generation a and b The chemically effective substructure fragments were randomly selected after being decomposed and identified by BRICS.

[0028] Chemical validity is verified through valence state checks and aromaticity corrections; molecules with more than 25 heavy atoms are automatically discarded.

[0029] Preferably, "mutation" reflects the heating and transformation process of the decoction, promoting chemical transformation and enhancing diversity; the mutation operators include functional group substitution, atomic substitution, bond insertion / deletion under valence rules, cyclization and aromatic rearrangement.

[0030] Preferably, in PPO reinforcement learning, each molecule is represented as a Markov decision process (S, A, P, R), where S represents the molecule state, A represents the edit action, P represents the transition probability, and R represents the reward function.

[0031] The molecular state S is encoded using a graph and embedded into the latent vector through a graph neural network encoder. The encoding formula is as follows:

[0032] ;

[0033] in, Let be the latent vector representation of the t-th molecule. The encoding function with φ as a parameter, The molecular diagram structure corresponding to the t-th molecule. A graph neural network model with φ as a parameter. It is an adjacency matrix. It is an atomic characteristic matrix;

[0034] The editing action A includes actions such as adding atoms, modifying bonds, or replacing functional groups;

[0035] Strategy Defines the state s in a given state t Choose action a t The probability of:

[0036]

[0037] Where softmax is the activation function. Here is the weight matrix for the action layer. The latent vector of the molecular state at step t. This is the bias term for the action layer.

[0038] Preferably, the reward function expression is:

[0039]

[0040] in, For the first t The reward value corresponding to the numerator. These are the weighting coefficients for each item. This is a penalty term for invalid structures.

[0041] Preferably, the iterative refinement termination condition includes the EGA evolution process termination condition and the PPO convergence condition;

[0042] The expression for the termination condition of the EGA evolution process is:

[0043] ;

[0044] in, The variance of fitness among molecular populations across generations. The number of molecules involved in the evaluation No. i Fitness value of each molecule This represents the average fitness of all molecules within the population. The preset EGA population convergence threshold is used;

[0045] The expression for the PPO convergence condition is:

[0046] ;

[0047] in, Let be the cumulative reward moving average of the numerator in generation t+1. Let be the cumulative reward moving average of the molecule in generation t. Preset convergence threshold for PPO strategy.

[0048] In summary, compared with traditional technologies, the beneficial effects of the present invention's method for identifying effective small molecules in traditional Chinese medicine prescriptions, which integrates EGA and PPO algorithms, are as follows:

[0049] (1) The molecules generated by the three algorithms are more potent and chemically effective than the original TPACD components. Among them, the BMGA module achieved the highest average affinity, providing a rich pool of highly active molecular skeletons for subsequent optimization; the EMGA module significantly improved bioaffinity and exploration efficiency, with the lowest structure alarm level, achieving the optimal balance between potency and chemical effectiveness; the IMGA module had the best overall performance, the highest drug-likeness, more compact molecular structure, moderate flexibility, the best docking stability, and the strongest predicted CRC inhibition potential.

[0050] (2) Through the closed-loop feedback mechanism of EGA and PPO, as well as the parallel collaboration and information interaction of the three modules, an adaptive discovery ecosystem was constructed, which avoids the blindness of traditional screening methods and greatly shortens the discovery cycle of active molecules.

[0051] (3) This invention establishes an interpretable analogy between traditional herbal processing procedures and evolutionary computation, providing a new approach for the modern research of traditional Chinese medicine compound formulas. At the same time, the algorithm parameters are clearly defined, possessing good repeatability and a standardized comparison basis, and can be extended to the mining of active ingredients in other traditional Chinese medicine compound formulas. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the molecular genetic algorithm for identifying effective small molecules in traditional Chinese medicine prescriptions using the EGA and PPO algorithms of this invention, specifically for the steaming and decocting process of traditional Chinese medicine.

[0053] Figure 2 This is a framework diagram of the intelligent molecular discovery system integrating BMGA, EMGA, and IMGA for identifying effective small molecules in traditional Chinese medicine prescriptions, which combines EGA and PPO algorithms according to the present invention.

[0054] Figure 3 This is a comparison chart of the affinity and drug-likeness characteristics of molecules generated by different algorithms for identifying effective small molecules in traditional Chinese medicine prescriptions, which integrates EGA and PPO algorithms according to this invention.

[0055] Among them, (a) is a histogram of affinity distribution of molecules generated by different algorithms, (b) is a scatter plot of the relationship between predicted affinity and drug-likeness score of molecules generated by different algorithms, and (c) is a violin distribution comparison plot of affinity of molecules generated by different algorithms.

[0056] Figure 4 This is a three-dimensional relationship diagram of molecular weight, LogP, and predicted affinity for effective small molecules in traditional Chinese medicine prescriptions, based on the integration of EGA and PPO algorithms in this invention.

[0057] Figure 5 This is a multi-dimensional comparison chart of the physicochemical properties and structural features of molecules generated by different algorithms for identifying effective small molecules in traditional Chinese medicine prescriptions, which integrates EGA and PPO algorithms according to this invention.

[0058] Among them, (a) is a scatter plot of molecular LogP and molecular weight (MW), (b) is a multi-algorithm distribution histogram of molecular weight (MW), topological polar surface area (TPSA) and LogP, and (c) is a distribution plot of rotatable bond violin.

[0059] Figure 6 This is a comparison diagram of the distribution of structural motifs and hydrogen bond characteristics of molecules generated by different algorithms for identifying effective small molecules in traditional Chinese medicine prescriptions using the EGA and PPO algorithms of this invention.

[0060] Among them, (a) is a box plot of the number of heavy atoms, (b) is a box plot of the number of rings, and (c) is a multi-algorithm distribution histogram of hydrogen bond acceptors (HBA) and hydrogen bond donors (HBD);

[0061] Figure 7 This is a comparison chart of the multi-attribute comprehensive performance of molecules generated by different algorithms for identifying effective small molecules in traditional Chinese medicine prescriptions, which integrates EGA and PPO algorithms according to this invention.

[0062] Among them, (a) is a multi-dimensional radar evaluation map, (b) is a comparison map of the radial stacking distribution of molecular core attributes, and (c) is a map of the differences in petal-shaped distribution characteristics.

[0063] Figure 8 This is a multi-dimensional comparison of the rule compliance and physicochemical properties of molecules in different algorithms for identifying effective small molecules in traditional Chinese medicine prescriptions, which integrates EGA and PPO algorithms according to this invention.

[0064] Among them, (a) is a radar chart of Lipinski / Veber rule violations, (b) is a three-dimensional waterfall plot of molecular physicochemical properties, and (c) is a planar heatmap of molecular core properties.

[0065] Figure 9 This invention generates a statistical comparison chart of the correlation features between molecular drug properties and molecular weight for different algorithms that integrate EGA and PPO algorithms to identify effective small molecules in traditional Chinese medicine prescriptions.

[0066] Among them, (a) is the QED violin distribution plot, (b) is the box plot of drug-likeness score, and (c) is the hexagonal density plot of molecular weight and drug-likeness.

[0067] Figure 10 This is a global correlation heatmap of key molecular attributes for identifying effective small molecules in traditional Chinese medicine prescriptions, based on the integration of EGA and PPO algorithms in this invention. Detailed Implementation

[0068] This invention proposes a method for identifying effective small molecules in traditional Chinese medicine prescriptions that integrates EGA and PPO algorithms. The technical method of this invention will be further described below with reference to the accompanying drawings and embodiments. The descriptions of the following embodiments are merely illustrative and are in no way intended to limit this application or its application or use.

[0069] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0070] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0071] Example 1

[0072] This embodiment presents the specific process of a method for identifying effective small molecules in traditional Chinese medicine prescriptions that integrates EGA and PPO algorithms.

[0073] First, the composition and proportion of the 18 Chinese herbs in Tiaopi Anchang Decoction (TPACD) were determined, as shown in Table 1. The known active ingredients of each Chinese herb were obtained, and phytochemicals with similar structures were collected to form an original compound collection containing more than 1,000 compounds.

[0074] Table 1

[0075]

[0076] To simulate the soaking stage of the decoction, the molecules need to be standardized and protonated, and simple structures need to be excluded through complexity screening to construct a molecular seed library.

[0077] Molecules are extracted from the molecular seed bank and initialized to form an initial population; the initialization scheme must satisfy the following: ;

[0078] Where mean represents the average value, MW is the molecular weight, and TPSA is the topological polar surface area. That is, the average molecular weight (MW) of all molecules in the initial population needs to be between 250 and 600, and the average topological polar surface area (TPSA) of all molecules in the initial population needs to be between 40 and 120.

[0079] The initial population is input into a computational system that integrates the Evolutionary Genetic Algorithm (EGA) and Proximal Policy Optimization (PPO) reinforcement learning. Through continuous molecular evolution, targeted screening, iterative refinement, and closed-loop feedback, a set of candidate molecules that meet the requirements is obtained. EGA performs large-scale molecular evolution, and PPO-based reinforcement learning enables adaptive fine-tuning. EGA simulates the diversity generation mechanism of herbal decoctions, while PPO mimics the adaptive reasoning process of physicians in prescription optimization, gradually improving molecular quality and predicting efficacy.

[0080] The computing system consists of three parallel and mutually supportive algorithm modules: BMGA (Basic Multi-Objective Genetic Algorithm), EMGA (Enhanced Multi-Objective Genetic Algorithm), and IMGA (Intelligent Multi-Objective Genetic Algorithm). The modules focus on different computational advantages rather than forming strict sequence relationships.

[0081] BMGA enables rapid, large-scale exploration through simplified mutation and crossover operators, generating molecular skeletons with diverse structures.

[0082] EMGA integrates affinity and drug-likeness scores to introduce predictive guidance, directing evolutionary processes toward biologically relevant regions in the chemical space.

[0083] IMGA employs a PPO-based strategy learning approach to iteratively refine the best-performing molecules, achieving multi-objective optimization of affinity, stability, and novelty.

[0084] These modules exchange information bidirectionally: the molecular diversity of BMGA provides a reference for the targeted evolution of EMGA; the screening results of EMGA provide seed molecules for the fine-tuning process of IMGA; and the optimized molecules output by IMGA are cyclically fed back to enrich the generative library. This cyclical interaction constructs an adaptive discovery ecosystem, enabling continuous refinement and knowledge reinforcement. The structures of BMGA, EMGA, and IMGA are as follows: Figure 2 As shown.

[0085] The specific processes of a computing system include:

[0086] The BMGA module is activated, and the initial population is evolved through simplified mutation and crossover operators. After selection and convergence, BMGA output molecular skeletons with diverse structures are obtained. This stage focuses on the extensive exploration of molecular structures, providing a rich candidate basis for subsequent optimization.

[0087] The molecular skeleton output from BMGA is input into the EMGA module, and affinity prediction and drug-likeness scoring are introduced as guides. Selection and convergence are performed again, and the evolutionary process is directed to a chemically relevant region, forming a high-potential molecule set of EMGA.

[0088] From the EMGA high-potential molecule set, select molecules with good performance as seed molecules, input them into the IMGA module, and perform iterative refinement through PPO reinforcement learning to obtain the IMGA optimized molecule set.

[0089] The optimized IMGA molecular set is fed back into the BMGA module to enrich the EGA population. At the same time, the screening results of EMGA are dynamically fed back into IMGA. The reward weight of reinforcement learning is adjusted, and S401 to S404 are repeated until the iterative refinement stopping condition is met.

[0090] like Figure 1 As shown, EGA establishes an interpretable analogy between traditional herbal processing procedures and evolutionary computation. Just as decoctions progressively transform raw herbal materials through heating, compatibility, and purification, EGA iteratively transforms molecular populations through mutation, recombination, and selection. This approach not only ensures chemical validity and structural novelty but also provides a conceptual framework that combines modern optimization principles with traditional pharmacognosy reasoning. The algorithm transforms each molecule... mi Encoded as a molecular diagram Gi ( V , E ),in V Represents atoms,E Represents a covalent bond.

[0091] "Crossing" corresponds to the compatibility stage in herbal processing, where new functional relationships are created by combining components. The expression for the cross operation is:

[0092] ;

[0093] in, This is a graph of offspring molecules generated through crossover operations. This is a crossover operation function. , Molecular diagrams of the two parent molecules involved in the crossover. For fragment assembly functions, and From the father generation a and b The chemically effective substructure fragments were randomly selected after being decomposed and identified by BRICS.

[0094] Chemical validity is verified through valence state checks and aromaticity corrections; molecules with more than 25 heavy atoms are automatically discarded.

[0095] "Mutation" reflects the heating and transformation process of the decoction, promoting chemical transformation and enhancing diversity; the mutation operators include functional group substitution, atomic substitution, bond insertion / deletion under valence rules, cyclization and aromatic rearrangement.

[0096] The "selection" algorithm simulates the heat control characteristics of a soup, employing a dynamic tournament selection strategy to adjust the selection pressure across different generations. ρ , ρ It grows exponentially, gradually prioritizing individuals with higher fitness; early generations emphasize diversity, for low... ρ Later generations focused on convergence, for high... ρ ;

[0097] The formula for calculating the probability of selection is:

[0098] ;

[0099] in, No. t molecule The probability of being selected to participate in subsequent evolution, where p is the selection pressure parameter. Let be the fitness value of the t-th molecule. It is the sum of the fitness values ​​of all molecules in the current population raised to the power of p.

[0100] The evolutionary process of molecular populations follows a multi-objective optimization principle, expressed as:

[0101] ;

[0102] in, For the first t One molecule to be evaluated Overall score QED This is a quantitative estimate of the drug-like properties. SA Represents the accessibility of synthesis. To predict affinity, For the structural novelty based on the Tanimoto distance, Here are the weighting coefficients for each indicator, and the weights... After normalization, it satisfies .

[0103] After each generation of evolution, candidate molecules are filtered for chemical stability, novelty, and drug-likeness, and ineffective smiles, PAINS-positive, or highly toxic structures are penalized.

[0104] ;

[0105] in, Represents the score for violations. Control the intensity of punishment. This is the original score. This is the final score after penalty adjustments. This filtration mechanism mimics the process of removing residues from a decoction, ensuring that only compounds with pharmacological potential and synthetic feasibility are retained.

[0106] As the population converges, the mutation probability... It exhibits exponential decay:

[0107] ;

[0108] in, k Controlling the cooling rate is analogous to the gradual reduction of heat observed during the processing of herbs, where e is a natural constant and t is the generation number of evolution. The initial mutation probability, Let be the mutation probability of an individual in generation t.

[0109] When the intergenerational fitness variance falls below the convergence threshold, the evolutionary process terminates.

[0110] ;

[0111] in, The variance of fitness among molecular populations across generations. The number of molecules involved in the evaluation No. i The overall score of each molecule This represents the average score of all molecules within the population. This is the preset EGA population convergence threshold.

[0112] This result indicates that the population is stable, analogous to the final concentration stage of a decoction.

[0113] EGA mimics herbal transformation at the population level, while PPO controls fine-grained, sequential molecular editing processes.

[0114] In PPO reinforcement learning, each molecule is represented as a Markov decision process (S, A, P, R), where S represents the molecule state, A represents the edit action, P represents the transition probability, and R represents the reward function.

[0115] The molecular state S is encoded using a graph and embedded into the latent vector through a graph neural network encoder. The encoding formula is as follows:

[0116] ;

[0117] This encoding captures the local topology and global electron distribution required to achieve property prediction, where, Let be the latent vector representation of the t-th molecule. The encoding function with φ as a parameter, The molecular diagram structure corresponding to the t-th molecule. A graph neural network model with φ as a parameter. It is an adjacency matrix. It is an atomic feature matrix.

[0118] The editing action A includes actions such as adding atoms, modifying bonds, or replacing functional groups;

[0119] Strategy Defined in a given state Choose action a t The probability of:

[0120] ;

[0121] Where softmax is the activation function. Here is the weight matrix for the action layer. This is the bias term for the action layer.

[0122] Each action modifies the molecular diagram and transitions the system to the next state. .

[0123] The total reward incentivizes the generation of effective, active, and drug-like molecules. The total reward function is expressed as follows:

[0124]

[0125] in, For the firstt The reward value corresponding to the numerator. These are the weighting coefficients for each item. This is a penalty term for invalid structures. This design ensures a trade-off between affinity-driven optimization and structural novelty.

[0126] PPO employs an actor-critic structure, where the actor... Proposed molecular modification schemes, critics Estimated expected cumulative reward:

[0127] ;

[0128] in, For the dominant function, This is the discount factor.

[0129] Editing target stable training process:

[0130] ;

[0131] in, For the pruning loss function, Expectation of time step t The ratio of selection probabilities before and after the policy update. For the clipping function, It is the preset clipping factor.

[0132] Training is performed over multiple epochs using mini-batch sampling optimized by Adam. Entropy regularization encourages exploration.

[0133] ;

[0134] Total loss The calculation expression is:

[0135] ;

[0136] in, This is the weighting coefficient for value loss. For the value function loss, This is the entropy regularization term.

[0137] The expression for the PPO convergence condition is:

[0138] ;

[0139] in, Let be the cumulative reward moving average of the numerator in generation t+1. Let be the cumulative reward moving average of the molecule in generation t. Preset convergence threshold for PPO strategy.

[0140] PPO optimization results are periodically reintegrated into the EGA population as elite seeds, increasing evolutionary diversity and accelerating convergence. This cyclical interaction forms a closed-loop optimization system, where EGA ensures chemical diversity and PPO ensures pharmacological precision.

[0141] Example 2

[0142] This embodiment verifies the technical effect of the invention by performing multi-dimensional attribute detection and statistical analysis on the candidate molecules output by the three modules BMGA, EMGA, and IMGA, and then plotting them.

[0143] As shown in Table 2, the basic algorithm achieved the highest average affinity (4.980), but produced larger and more flexible molecules (molecular weight ≈ 307.5, rotatable bonds ≈ 3.29), and the lowest drug-likeness (QED) (0.528). The enhanced method produced structures with moderate to light mass (molecular weight ≈ 251.1), balanced affinity (4.619), and the lowest structural alarm level (0.022), indicating an optimal trade-off between potency and chemoeffectiveness. The intelligent method produced compounds with the best pharmacological properties, the highest drug-likeness (0.590), more compact structure (molecular weight ≈ 238.5), and moderate flexibility (rotatable bonds ≈ 2.61), although its affinity was slightly lower (4.495) and its structural alarm level was slightly higher (0.171).

[0144] The Basic Multi-Objective Genetic Algorithm (BMGA) excels in terms of initial binding strength, while the Enhanced Multi-Objective Genetic Algorithm (EMGA) and the Intelligent Multi-Objective Genetic Algorithm (IMGA) generate more refined and drug-like structures. These three models collectively outline the Pareto front between affinity-driven optimization and drug-likeness-driven optimization.

[0145] Table 2

[0146]

[0147] like Figure 3 As shown in (a), the performance of the basic algorithm (BMGA), the enhancement method (EMGA), and the intelligent method (IMGA) differs significantly in the task of generating anti-CRC leader molecules for the traditional Chinese medicine Tiaopi Anchang Decoction (TPACD). The basic algorithm has the widest distribution range and the highest overall bias, indicating that the molecules it generates tend to have stronger binding affinity. The enhancement method and the intelligent method have narrower distribution ranges, concentrated in the medium affinity range, which means that their optimization process is more controllable than random exploration.

[0148] like Figure 3As shown in (b), the enhancement and intelligent methods form dense clusters in the high drug-likeness region (drug-likeness > 0.7), while the basic method has a wider distribution, with some outliers exhibiting high affinity but low drug-likeness. These findings suggest that the improved structure filtering process in the EMGA / IMGA scheme reduces the number of extreme cases, indicating that pharmacologically significant candidate molecules are more favored. Figure 3 (c) This observation is further supported by the fact that the enhanced and intelligent methods have lower variances, reflecting the consistency and stability of their molecular weight.

[0149] Figure 4 A deeper level of structural correlation was revealed. Basic methods mainly achieve high affinity through high molecular weight and high LogP compounds, while enhanced and smart methods can still maintain satisfactory affinity in the low molecular weight to low LogP range, which is a key feature of oral bioavailability.

[0150] The study of the structure and physicochemical properties of the generated molecules, such as Figure 5 As shown in (a), the molecules generated by EMGA and IMGA are concentrated in the medium molecular weight and LogP region, with topological polar surface area (TPSA) between 50 and 70, indicating that their permeability-solubility balance is optimal. In contrast, the molecules generated by BMGA are biased towards higher molecular weight and the LogP region, which is consistent with its affinity-oriented properties.

[0151] like Figure 5 As shown in (b), the distribution of enhancement and intelligent methods shifts to the left: predominantly consisting of lighter, less polar compounds, consistent with the improved efficacy of orally administered drugs. Figure 5 (c) and Figure 6 As shown in (a), the skeletons generated by the enhancement method and the intelligent method are more rigid, easier to synthesize, have lower degrees of freedom, and stronger conformational stability. Figure 6 (c) shows that both the enhancement method and the intelligent method are in high agreement with Lipinski's five rules, confirming that they have good polarity and hydrogen bonding properties.

[0152] To assess the overall trend, Figure 7 Integrate multiple indicators into a composite visual chart. For example... Figure 7 As shown in (a), the radar chart shows that the enhanced and intelligent methods outperform the basic methods in terms of drug-likeness, rule compliance and structural compactness, while the basic methods only lead in terms of affinity. Figure 7 (b) and Figure 7 (c) This view is further reinforced: the maps of enhancement and intelligent methods are compact and balanced, while the petals of basic methods are more widely distributed, indicating that they are too molecularly massive and flexible.

[0153] Comparison of rule violations, such as Figure 8As shown in (a), the improvements from the augmentation and intelligent methods are significant, reducing the number of violations of the Lipinski / Veber rule by nearly two-thirds. However, the intelligent method has a slightly higher level of structural alarms, suggesting that additional PAINS filtering is needed in future optimizations. Figure 8 (b) and Figure 8 (c) It was confirmed that molecular weight, TPSA and flexibility showed a continuous downward trend, providing quantitative evidence for improving drug-likeness with minimal loss of affinity.

[0154] Figure 9 The supplementary visualization analysis further enhances the statistical robustness of the performance comparison results of different algorithms. Among them, Figure 9 The statistical characteristics in (b) show that the median drug-likeness score of molecules generated by the enhancement method and the intelligent method is significantly improved. This result confirms that the two types of advanced models not only improve the average drug-like potential of molecules, but also effectively reduce the overall variability of molecular properties, ensuring the uniformity of the quality of generated molecules. Figure 9 (c) clearly captures the negative correlation between molecular weight and drug-likeness, and the molecules generated by the enhancement method and the intelligent method are highly enriched in the low molecular weight-high drug-likeness quadrant, highlighting their technical superiority in balancing molecular size and drug potential.

[0155] at last, Figure 10 Quantitative analysis supports these explanations: molecular weight is positively correlated with TPSA (r > 0.7), LogP is negatively correlated with TPSA (r < -0.6), and drug-likeness is negatively correlated with rule violation (r < -0.8), which further confirms the balance achieved by EMGA and IMGA.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for identifying effective small molecules in traditional Chinese medicine prescriptions by integrating EGA and PPO algorithms, characterized in that, Includes the following steps: S1. Obtain the known components of the traditional Chinese medicine Tiaopi Anchang Decoction (TPACD) and phytochemicals with structures similar to the known components, forming an original compound set containing more than 1,000 compounds. S2. Standardize and protonate the molecules in the original compound set, and eliminate simple structures by complexity screening to construct a molecular seed library; S3. Extract molecules from the molecular seed bank and initialize them to form an initial population; S4. Input the initial population into a computational system that integrates the Evolutionary Genetic Algorithm (EGA) and the PPO reinforcement learning algorithm. Through continuous molecular evolution, targeted screening, iterative refinement, and closed-loop feedback, a set of candidate molecules that meet the requirements is obtained. The computing system includes the basic multi-objective genetic algorithm BMGA, the enhanced multi-objective genetic algorithm EMGA, and the intelligent multi-objective genetic algorithm IMGA. The specific content of S4 includes: S401. Start the BMGA module and evolve the initial population through simplified mutation and crossover operators. After selection and convergence, obtain the structurally diverse BMGA output molecular backbone. S402. Input the BMGA output molecular skeleton into the EMGA module, introduce affinity prediction and drug-likeness scoring as guidance, and perform selection and convergence again to guide the evolutionary process to a chemically relevant chemical spatial region to form an EMGA high-potential molecule set. S403. Select high-performing molecules from the EMGA high-potential molecule set as seed molecules, input them into the IMGA module, and perform iterative refinement through PPO reinforcement learning to obtain the IMGA optimized molecule set. S404. Feed the optimized IMGA molecular set back to the BMGA module to enrich the EGA population. At the same time, dynamically feed the screening results of EMGA back to IMGA, adjust the reward weight of reinforcement learning, and repeat S401 to S404 until the iterative refinement stopping condition is met. In each generation of evolution, candidate molecules are filtered for chemical stability, novelty, and drug-likeness, and ineffective SMILES, PAINS positive, or highly toxic structures are penalized.

2. The method for identifying effective small molecules in traditional Chinese medicine prescriptions by integrating EGA and PPO algorithms according to claim 1, characterized in that, The evolutionary process of molecular populations follows a multi-objective optimization principle, expressed as: ； in, For the first t One molecule to be evaluated Overall score QED This is a quantitative estimate of the drug-like properties. SA Represents the accessibility of synthesis. To predict affinity, For the structural novelty based on the Tanimoto distance, Here are the weighting coefficients for each indicator, and the weights... After normalization, it satisfies .

3. The method for identifying effective small molecules in traditional Chinese medicine prescriptions by integrating EGA and PPO algorithms according to claim 1, characterized in that, The EGA algorithm employs a dynamic tournament selection strategy, adjusting the selection pressure across different generations. ρ , ρ It grows exponentially, gradually prioritizing individuals with higher fitness; early generations emphasize diversity, for low... ρ Later generations focused on convergence, for high... ρ The formula for calculating the probability of selection is: ； in, No. t molecule The probability of being selected to participate in subsequent evolution, where p is the selection pressure parameter. Let be the fitness value of the t-th molecule. It is the sum of the fitness values ​​of all molecules in the current population raised to the power of p.

4. The method for identifying effective small molecules in traditional Chinese medicine prescriptions by integrating EGA and PPO algorithms according to claim 1, characterized in that, The expression for the crossover operation is: ； in, This is a graph of offspring molecules generated through crossover operations. This is a crossover operation function. , Molecular diagrams of the two parent molecules involved in the crossover. For fragment assembly functions, and From the father generation a and b The chemically effective substructure fragments were randomly selected after being decomposed and identified by BRICS. Chemical validity is verified through valence state checks and aromaticity corrections; molecules with more than 25 heavy atoms are automatically discarded.

5. The method for identifying effective small molecules in traditional Chinese medicine prescriptions by integrating EGA and PPO algorithms according to claim 1, characterized in that, "Mutation" reflects the heating and transformation process of the decoction, promoting chemical transformation and enhancing diversity; the mutation operators include functional group substitution, atomic substitution, bond insertion / deletion under valence rules, cyclization and aromatic rearrangement.

6. The method for identifying effective small molecules in traditional Chinese medicine prescriptions by integrating EGA and PPO algorithms according to claim 1, characterized in that, In PPO reinforcement learning, each molecule is represented as a Markov decision process (S, A, P, R), where S represents the molecule state, A represents the edit action, P represents the transition probability, and R represents the reward function. The molecular state S is encoded using a graph and embedded into the latent vector through a graph neural network encoder. The encoding formula is as follows: ； in, Let be the latent vector representation of the t-th molecule. The encoding function with φ as a parameter, The molecular diagram structure corresponding to the t-th molecule. A graph neural network model with φ as a parameter. It is an adjacency matrix. It is the atomic characteristic matrix; The editing action A includes actions such as adding atoms, modifying bonds, or replacing functional groups; Strategy Defines the state s in a given state t Choose action a t The probability is expressed as: ； Where softmax is the activation function. Here is the weight matrix for the action layer. The latent vector of the molecular state at step t. This is the bias term for the action layer.

7. The method for identifying effective small molecules in traditional Chinese medicine prescriptions by integrating EGA and PPO algorithms according to claim 6, characterized in that, The reward function expression is: ； in, For the first t The reward value corresponding to each molecule. These are the weighting coefficients for each item. This is a penalty term for invalid structures.

8. The method for identifying effective small molecules in traditional Chinese medicine prescriptions by integrating EGA and PPO algorithms according to claim 1, characterized in that, The iterative refinement stopping conditions include the EGA evolution process termination condition and the PPO convergence condition; The expression for the termination condition of the EGA evolution process is: ； in, The variance of fitness among molecular populations across generations. The number of molecules involved in the evaluation No. i Fitness value of each molecule This represents the average fitness of all molecules within the population. The preset EGA population convergence threshold is used; The expression for the PPO convergence condition is: ； in, Let be the cumulative reward moving average of the numerator in generation t+1. Let be the cumulative reward moving average of the molecule in generation t. Preset convergence threshold for PPO strategy.

Images