A method and system for generating drug molecules by reinforced backbone clustering

Abstract

The application discloses a method and system for generating drug molecules by reinforcing skeleton clustering, and the method comprises the following steps: preprocessing a molecular data set and pre-training a sequence generation model to obtain an initial strategy network; taking the initial strategy network as a current strategy network to iteratively execute: generating a candidate molecule, extracting a core skeleton, and clustering the core skeleton into multiple skeleton clusters based on similarity; performing multidimensional scoring on each candidate molecule based on the skeleton clusters in terms of biological activity, structural diversity and novelty of the skeleton clusters, identifying an optimal molecule set through a Pareto frontier analysis, and calculating a composite reward value; updating the parameters of the strategy network in the direction of the composite reward value until convergence to obtain an optimized strategy network, and generating a final drug candidate molecule from the optimized strategy network. The application deeply integrates skeleton clustering and Pareto multi-objective evaluation into a reinforcement learning decision cycle, and solves the problems of serious skeleton homogenization, difficult multi-objective trade-off and low exploration efficiency of existing methods.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and drug design technology, and in particular to a method and system for generating drug molecules using enhanced scaffold clustering. Background Technology

[0002] One of the core tasks of drug discovery is to efficiently design novel molecular structures that possess ideal biological activity, excellent drug-like properties, and synthetic feasibility. This is also the most critical upstream link in the entire process of innovative drug development. For a long time, traditional experimental-based high-throughput screening methods have been the mainstream approach for lead compound discovery. However, this method not only has a long development cycle, but its explorable molecular range is also limited by existing physical compound libraries, making it difficult to cover a broad chemical space, which greatly restricts the discovery efficiency and structural diversity of novel lead compounds.

[0003] In recent years, with the rapid development of artificial intelligence and computational chemistry, deep learning-based molecular generation methods have gradually become a cutting-edge direction in drug development, providing a new technological path to overcome the bottlenecks of traditional drug discovery. Early computational molecular design methods mainly relied on template splicing or fragment assembly based on chemical rules. The molecular structures they could generate were limited by preset templates, resulting in limited chemical space coverage and insufficient structural innovation. Deep generative models, represented by variational autoencoders (VAEs), generative adversarial networks (GANs), recurrent neural networks based on SMILES sequences, and Transformer autoregressive models, can learn the intrinsic distribution patterns of chemical structures from massive amounts of known molecular data, autonomously generating grammatically valid and chemically sound new molecular structures, significantly expanding the explorable chemical space. However, these data-driven generative methods primarily aim to reproduce the molecular distribution of the training set, and still have significant limitations in core scenarios such as multi-objective directional optimization and satisfying multiple drug-making constraints required for drug molecule design.

[0004] To address the insufficient targeted optimization capabilities of deep generative models, researchers have introduced reinforcement learning frameworks into the field of molecular generation, modeling the generation process of molecular sequences as a sequential decision-making task. Within this framework, the generative model acts as an agent; each step of its sequence generation action receives a reward based on the predicted drug-grade properties of the final generated molecule. The model parameters are then iteratively updated using a policy gradient algorithm based on the reward signal. This design significantly enhances the generative model's targeted optimization capabilities for target properties. However, existing reinforcement learning-based drug molecule generation techniques still suffer from significant technical limitations in practical applications.

[0005] On the one hand, the model is prone to getting stuck in local optima during the optimization process, tending to repeatedly generate molecular variants with similar core skeletons. This results in severe homogenization of the generated skeletons, low efficiency in chemical space exploration, and difficulty in discovering novel lead compounds. On the other hand, most existing methods perform generation operations at the atomic or subgraph level, lacking explicit planning and guidance for the skeleton structure that determines the core physicochemical properties and biological activities of molecules. The generation process is highly blind and easily produces ineffective molecules with complex synthetic routes and low drug potential. At the same time, existing methods sacrifice structural diversity in pursuit of target attribute scores, or significantly reduce the efficiency of target optimization in encouraging diversity, failing to meet the core requirements of both.

[0006] To address the issue of insufficient molecular diversity, existing technologies have introduced improved schemes that incorporate cluster analysis. These include clustering molecules during training data preprocessing or introducing a diversity penalty term into the model's loss function. However, most of these schemes are loosely coupled with the reinforcement learning's generation and optimization process, existing only as data preprocessing steps or auxiliary constraints. They fail to deeply integrate the structural information and cluster distribution characteristics of the molecular skeleton into the core decision-making loop of reinforcement learning, thus failing to form a systematic and hierarchical generation guidance mechanism. Consequently, they cannot fundamentally solve the core problems of skeleton homogeneity and low exploration efficiency. Summary of the Invention

[0007] Therefore, the technical problem to be solved by the present invention is to overcome the problem in the prior art that the lack of explicit guidance and cluster analysis of molecular skeleton leads to the generation process getting stuck in local optima, serious homogenization of skeleton, and difficulty in taking into account both biological activity and structural novelty in multi-objective optimization.

[0008] To address the aforementioned technical problems, this invention provides a method for generating drug molecules using enhanced backbone clustering, comprising: The molecular dataset is preprocessed, and the pre-trained sequences of the preprocessed molecular dataset are used to generate a model, resulting in an initial policy network. The initial policy network is used as the current policy network for the first round of iteration, and candidate molecules are generated from the current policy network. The core backbone of each candidate molecule is extracted, and the candidate molecules are clustered based on backbone similarity to form multiple backbone clusters. Based on the aforementioned skeletal cluster, each candidate molecule is scored using multi-dimensional attributes, including bioactivity score, structural diversity score, and skeletal cluster novelty bonus. Pareto front analysis is performed based on the multi-dimensional scores of all candidate molecules to identify the Pareto optimal molecular set. A composite reward value is calculated for each candidate molecule based on its multi-dimensional score and whether it belongs to the Pareto optimal molecular set. Guided by the composite reward value, the parameters of the policy network are updated, and the updated policy network is used as the current policy network for the next iteration until the performance of the policy network converges or the preset number of iterations is reached, thus obtaining the optimized policy network. The candidate molecules generated by the optimization strategy network are used as the final output drug candidate molecules.

[0009] In one embodiment of the present invention, the method for obtaining the initial policy network by pre-training a sequence generation model using a preprocessed molecular dataset is as follows: A neural network based on a self-attention mechanism is constructed as the sequence generation model. The simplified linear input canonical sequences from the preprocessed molecular dataset are used as the training dataset. Maximum likelihood estimation is employed as the pre-training objective, maximizing the conditional log-likelihood of the sequence. The mathematical expression is: , in, This represents a simplified linear input canonical sequence of a molecule, where T is the sequence length. For the t-th character, This represents the historical sequence preceding the t-th character. For policy networks, For model parameters, The training dataset is used; the model parameters are iteratively updated, and after training, the learned model parameters are used as the initial policy network.

[0010] In one embodiment of the present invention, the method for extracting the core skeleton of each candidate molecule is as follows: the simplified linear input canonical representation of each molecule is analyzed to identify all ring systems in the molecule, the ring systems and the atoms and bonds on the shortest path connecting the ring systems are retained, and all non-cyclic side chain atoms and substituents are removed to obtain a simplified structure containing only ring systems and linkers, which serves as the core skeleton of the molecule.

[0011] In one embodiment of the present invention, the method for clustering the candidate molecules into multiple backbone clusters based on backbone similarity is as follows: the core backbone of each molecule is converted into a molecular fingerprint, and the similarity between any two backbones is calculated. and The formula for the Tanimoto similarity between them is: , in, This represents the fingerprint generation function. and Let represent the intersection and union of the sets respectively; construct a similarity matrix based on the similarity of all molecular pairs, and divide the molecular set into multiple skeleton clusters through unsupervised clustering, with each cluster representing a chemical subspace with similar core structures.

[0012] In one embodiment of the present invention, based on the scaffold cluster, each candidate molecule is scored with multi-dimensional attributes. The multi-dimensional attribute scores include biological activity score, structural diversity score and scaffold cluster novelty reward. Specifically, it includes: hierarchically sorting each scaffold cluster according to the Pareto non-dominated level of the candidate molecules in each scaffold cluster, constructing a hierarchical molecular cluster structure with different priorities, and different levels corresponding to different intra-cluster reward regulation coefficients. Based on the hierarchical molecular cluster structure, each candidate molecule is scored with multi-dimensional attributes, including bioactivity score, structural diversity score, and backbone cluster novelty reward. The backbone cluster novelty reward is dynamically calculated based on the hierarchical position of the cluster to which the molecule belongs and the sparsity within the cluster.

[0013] In one embodiment of the present invention, the method for identifying the Pareto optimal molecular set by performing Pareto front analysis based on the multi-dimensional scores of all candidate molecules is as follows: The bioactivity score, structural diversity score, and scaffold cluster novelty reward of each candidate molecule are constructed as a three-dimensional attribute feature vector, mapping each candidate molecule to a data point in the three-dimensional attribute space; the dominance relationship between molecules is defined: for any two candidate molecules M1 and M2, if M1's scores in all three dimensions (bioactivity score, structural diversity score, and scaffold cluster novelty reward) are not lower than M2's, and M1's score in at least one dimension is strictly higher than M2's, then M1 is determined to dominate M2, and M2 is a dominated molecule; all candidate molecules in the current batch are sorted for non-dominated status, and all non-dominated molecules not dominated by any other candidate molecules are selected. The set of non-dominated molecules is identified as the first Pareto front, serving as the Pareto optimal molecular set.

[0014] In one embodiment of the present invention, the method for calculating the composite reward value for each candidate molecule M is as follows: , in, To score bioactivity, To score structural diversity, As a reward for the novelty of the skeleton cluster, , , The preset weighting coefficients satisfy... ; This is an indicator function; its value is 1 when the molecule is at the Pareto front, and 0 otherwise. It is the Pareto optimal molecular set; This refers to the frontier reward intensity coefficient.

[0015] In one embodiment of the present invention, the method for updating the parameters of the policy network with the composite reward value as the optimization guide is as follows: Reinforcement learning algorithm is used to update the strategy network parameters. The objective function is: , in, For probability ratios, This is the estimated value of the dominance function. For the clipping function, These are the trimming parameters; By maximizing this objective function, the policy network is guided to generate drug candidate molecules with high composite reward values ​​while ensuring training stability.

[0016] In one embodiment of the present invention, the method for preprocessing the molecular dataset is as follows: performing syntax checks and valence validity verification on the simplified linear input canonical representation of each molecule, and removing invalid molecules; converting the valid simplified linear input canonical representation of the molecules into a standard canonical form to ensure uniqueness; removing duplicate molecules to obtain the preprocessed training dataset.

[0017] Based on the same inventive concept, the present invention also provides a drug molecule generation system with enhanced backbone clustering, comprising the following modules: The data preprocessing and model pretraining module is used to preprocess the molecular dataset, generate a model using the pre-trained sequence of the preprocessed molecular dataset, and obtain the initial policy network. The iterative optimization module is used to take the initial policy network as the current policy network for the first round of iteration, generate candidate molecules from the current policy network, extract the core backbone of each candidate molecule, and cluster the candidate molecules based on backbone similarity to divide them into multiple backbone clusters. Based on the scaffold cluster, each candidate molecule is scored with multi-dimensional attributes, including bioactivity score, structural diversity score, and scaffold cluster novelty bonus. Pareto front analysis is performed based on the multi-dimensional scores of all candidate molecules to identify the Pareto optimal molecular set. The composite reward value is calculated based on the multi-dimensional scores of each candidate molecule and whether it belongs to the Pareto optimal molecular set. Guided by the composite reward value, the parameters of the policy network are updated, and the updated policy network is used as the current policy network for the next iteration until the performance of the policy network converges or the preset number of iterations is reached, thus obtaining the optimized policy network. The result output module is used to take the candidate molecules generated by the optimization strategy network as the final output drug candidate molecules.

[0018] The technical solution of the present invention has the following advantages compared with the prior art: This invention constructs a Pareto front analysis based on three dimensions—bioactivity, structural diversity, and skeletal cluster novelty—and calculates a composite reward based on this, achieving an automatic optimal trade-off between multiple objective attributes. It introduces molecular skeletal-based cluster analysis, transforming chemical space into a structured skeletal cluster representation, and designs a novelty reward mechanism inversely proportional to cluster size. This overcomes the problem of skeletal homogenization in existing methods, improving the efficiency of chemical space exploration and the probability of discovering novel lead compounds. The molecules generated by this invention outperform existing mainstream methods in predicting bioactivity, drug-likeness, and structural novelty, demonstrating outstanding overall quality and practicality. Attached Figure Description

[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0020] Figure 1 This is a schematic flowchart of the enhanced backbone clustering drug molecule generation method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the molecular generation method framework based on reinforcement learning and molecular backbone clustering in an embodiment of the present invention; Figure 3 yes Figure 2 By capturing atomic-level information of molecules, a magnified schematic diagram of the molecular core framework is obtained. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0022] Example 1: like Figure 1 As shown, this invention provides a method for generating drug molecules using enhanced backbone clustering, comprising: The molecular dataset is preprocessed, and the pre-trained sequences of the preprocessed molecular dataset are used to generate a model, resulting in an initial policy network. The initial policy network is used as the current policy network for the first round of iteration, and candidate molecules are generated from the current policy network. The core backbone of each candidate molecule is extracted, and the candidate molecules are clustered based on backbone similarity to form multiple backbone clusters. Based on the scaffold cluster, each candidate molecule is scored with multi-dimensional attributes, including bioactivity score, structural diversity score, and scaffold cluster novelty bonus. Pareto front analysis is performed based on the multi-dimensional scores of all candidate molecules to identify the Pareto optimal molecular set. The composite reward value is calculated based on the multi-dimensional scores of each candidate molecule and whether it belongs to the Pareto optimal molecular set. Guided by the composite reward value, the parameters of the policy network are updated, and the updated policy network is used as the current policy network for the next iteration until the performance of the policy network converges or the preset number of iterations is reached, thus obtaining the optimized policy network. The candidate molecules generated by the optimization strategy network are used as the final output drug candidate molecules.

[0023] This invention preprocesses molecular datasets and pre-trains sequence generation models to obtain an initial policy network. It then deeply integrates reinforcement learning with backbone clustering analysis through iterative optimization: in each iteration, the current policy network generates candidate molecules, extracts their core backbones, and clusters them to construct a structured chemical subspace representation. Based on this, each molecule is scored across multiple dimensions, including bioactivity, structural diversity, and backbone cluster novelty. Pareto front analysis identifies the optimal molecule set and calculates a composite reward value, which guides the updating of the policy network parameters until convergence. By introducing backbone clustering and Pareto multi-objective optimization mechanisms, this invention systematically addresses the problems of severe backbone homogenization, difficulty in balancing multiple objectives, and low exploration efficiency in existing methods. It achieves explicit guidance of the molecular backbone during reinforcement learning optimization, significantly improving the breadth and depth of chemical space exploration. Furthermore, its modular design provides good versatility and scalability, offering reliable technical support for the efficient generation of drug candidate molecules with both bioactivity and structural novelty.

[0024] In this embodiment of the invention, the input molecular dataset is first standardized and preprocessed. The purpose of the preprocessing step is to provide a high-quality, standardized molecular data foundation for subsequent model training, ensuring that the sequence generation model can learn the correct chemical syntax and structural distribution.

[0025] Specifically, preprocessing includes: using cheminformatics tools to perform syntax checks and valence validation on the simplified linear input canonical representation of each molecule, removing molecules with invalid syntax or unreasonable valences, as these molecules have no practical chemical significance and their inclusion in the training set would interfere with the model's correct learning of valid chemical structures. Valid simplified linear input canonical representations are then converted into standard canonical forms. For example, cheminformatics tools such as RDKit are used to unify different representations of the same molecule into a canonical string, ensuring that each molecule has a unique representation and preventing the model from misclassifying the same molecule as different samples due to differences in representation. Duplicate molecules are then removed, resulting in the preprocessed training dataset. After preprocessing, each molecule in the training dataset has a correct chemical structure, a unified representation, and uniqueness.

[0026] Furthermore, after preprocessing, a sequence generation model is pre-trained using the processed dataset as the initial policy network for the reinforcement learning agent. Sequence generation models are an important class of methods in deep generative models, transforming the molecular generation problem into a sequence prediction problem, that is, predicting the next character sequentially based on the existing characters, and finally generating a complete simplified linear input canonical sequence of molecules.

[0027] Unlike traditional rule-based or template-based molecular generation methods, sequence generation models can automatically learn the statistical regularities of chemical structures from large amounts of molecular data without requiring manual setting of generation rules. This invention constructs a neural network based on a self-attention mechanism as the sequence generation model, using Simplified Molecular Linear Input Canonical (SMILES) sequences from a preprocessed molecular dataset as training samples. The self-attention mechanism is the core component of the sequence generation model, allowing the model to directly focus on information from all other positions in the sequence when processing a particular position, thereby capturing long-distance dependencies. Compared to traditional recurrent neural networks, the self-attention mechanism better preserves global contextual information when processing long sequences and has higher computational efficiency, showing significant advantages in processing SMILES sequences with large length variations and complex structures. Furthermore, to further enhance the model's ability to perceive the local chemical environment and detailed structural features of molecules, this invention also designs an unfolded shadow feature extraction module in the sequence generation model. This module directly performs dimensional unfolding and linear projection transformation on the feature sequence based on the self-attention feature map, extracting shadow feature representations at different semantic levels.

[0028] Specifically, the feature vector output by the self-attention layer is expanded into a sliding window along the sequence length dimension and mapped to the shadow feature space through a learnable projection matrix, thereby obtaining the contextual substructure information around each position. Expanding the shadow features can effectively capture detailed information such as local bonding patterns between atoms and spatial arrangement of functional groups, and residual fusion is performed with the original self-attention features, thereby improving the generative model's ability to represent molecular topology and generating more chemically plausible and diverse molecular sequences.

[0029] In the pre-training phase, maximum likelihood estimation is used as the optimization objective, aiming to maximize the conditional log-likelihood of all sequences in the training dataset. Maximum likelihood estimation is one of the most commonly used parameter estimation methods in statistical learning. Its basic idea is to select a set of model parameters that maximizes the probability of observed training data occurrences. In the sequence generation task, this optimization objective aims to enable the sequence generation model to accurately predict every character of every molecular sequence in the training set. The mathematical expression of the objective function is: , in, This represents a simplified linear input canonical sequence of a molecule, where T is the sequence length. For the t-th character, This represents the historical sequence preceding the t-th character. For policy networks, For model parameters, This serves as the training dataset. By maximizing this objective function, the model learns the grammatical rules, chemical valence constraints, and molecular structure distribution patterns that simplify linear input canonical sequences, thereby mastering the ability to generate grammatically valid and chemically sound molecular sequences.

[0030] The objective function means that for each molecular sequence in the training set, the sequence generation model needs to accurately predict the character at the current position based on the historical characters preceding each position, maximizing the logarithmic sum of the predicted probabilities at all positions. The logarithmic operation transforms the product form into a summation form, simplifying the calculation and avoiding the numerical underflow problem that can occur when multiplying multiple probabilities. The expectation symbol represents the average of all sequences in the training dataset, which is approximated during actual training by randomly sampling batches of data.

[0031] By maximizing this expected value, sequence generation models can learn the basic grammatical rules for simplifying linear molecular input, chemical valence constraints, and the distribution patterns of common molecular structures, thereby mastering the ability to generate grammatically valid and chemically sound molecular sequences. For example, sequence generation models need to learn that the closing symbols of rings must appear in pairs, the connections between atoms must conform to the valence rules, and the common connection methods of different substituents, etc. This knowledge is acquired autonomously by the sequence generation model through a large amount of data. During training, the model parameters are iteratively updated using a gradient descent-type optimizer, and an appropriate learning rate scheduling strategy is adopted until the model's loss on the validation set converges or the preset number of training epochs is reached.

[0032] After pre-training, the learned model parameters are used as the initial policy network, providing a solid foundation of prior knowledge in chemical language for the subsequent reinforcement learning stages. Pre-training provides a reasonable starting point for the subsequent reinforcement learning optimization stages, avoiding the policy network from exploring from random initialization, thereby improving optimization efficiency and generation quality.

[0033] Furthermore, after obtaining the initial policy network, the initial policy network is used as the current policy network for the first round of iteration. A batch of candidate molecules is generated by sampling from the current policy network. The generation process adopts an autoregressive approach. The policy network predicts the next character step by step based on the generated character sequence until a complete simplified molecular linear input canonical sequence is generated, thereby obtaining a batch of candidate molecules with diverse structures.

[0034] like Figure 2 and Figure 3As shown, further, the core skeleton of each generated candidate molecule is extracted. In medicinal chemistry, the core skeleton of a molecule typically refers to the ring system skeleton that retains the basic structural features of the molecule, determining the overall configuration and key physicochemical properties of the molecule, while side chains and substituents mainly affect the subtle pharmacokinetic regulation of the molecule. Most existing molecular generation methods operate at the atomic or subgraph level, lacking explicit modeling of this higher-level structural feature, resulting in a certain degree of blindness in the generation process.

[0035] To overcome this problem, in this embodiment of the invention, the simplified linear input canonical representation of each molecule is analyzed to identify all ring systems in the molecule. The ring systems and the atoms and bonds on the shortest paths connecting them are retained, while all non-cyclic side-chain atoms and substituents are removed, resulting in a simplified structure containing only ring systems and linkers, which serves as the core skeleton of the molecule. This process strips away the variable parts of the molecular structure, retaining the core nucleus that determines the essential characteristics of the molecule, providing a stable and chemically meaningful structural basis for subsequent skeleton similarity analysis and clustering.

[0036] Furthermore, after extracting the core skeleton, the candidate molecules are clustered based on the Tanimoto similarity between the skeletons to form multiple molecular clusters.

[0037] Furthermore, a three-dimensional attribute feature vector is calculated for each candidate molecule. This feature vector includes a bioactivity score, a structural diversity score, and a skeletal cluster novelty reward. Based on these scores, all candidate molecules in the current batch are ranked according to Pareto non-dominated status to determine the Pareto non-dominated level of each molecule. According to the Pareto non-dominated level distribution of candidate molecules within each molecular cluster, the molecular clusters are hierarchically ranked: clusters containing a higher proportion of high-level non-dominated molecules (such as first Pareto front molecules) are placed at the upper level of the hierarchy, while clusters with lower priority (such as those with a majority of dominated molecules) are placed at the lower level, thus constructing a hierarchical molecular cluster structure with different priorities. Different levels correspond to different intra-cluster reward adjustment coefficients, with higher-level molecular clusters enjoying higher reward coefficients.

[0038] Based on the Pareto front analysis results (i.e., the non-dominated levels of multi-objective attribute scores) of candidate molecules within each molecular cluster, the molecular clusters are hierarchically ranked, with clusters of higher Pareto priority placed at the upper level and clusters of lower priority at the lower level, thus constructing a hierarchical molecular cluster structure with different priorities. Different levels correspond to different intra-cluster reward adjustment coefficients, with higher-level molecular clusters enjoying higher reward coefficients. Based on the hierarchical molecular clusters, each candidate molecule is scored for multi-dimensional attributes, including bioactivity score, structural diversity score, and skeletal cluster novelty reward, where the novelty reward is dynamically calculated based on the hierarchical position of the molecule within its cluster and the sparsity within the cluster. Then, Pareto front analysis is performed based on the multi-dimensional scores of all candidate molecules to identify the Pareto optimal molecular set. Finally, a composite reward value is calculated for each candidate molecule based on its multi-dimensional score and whether it belongs to the Pareto optimal molecular set.

[0039] Specifically, the simplified linear input canonical sequence of each core skeleton is input into a pre-constructed bidirectional long short-term memory network (bidirectional LSTM). This network captures the contextual dependencies of atoms and chemical bonds in the sequence from both forward and backward directions, and extracts deep feature vectors containing information on skeleton topology and ring system connections.

[0040] The high-dimensional feature sequence output from the bidirectional LSTM is linearly transformed through a projection layer to reduce its dimensionality to a low-dimensional continuous space. Then, a flattening operation aggregates the sequence features into a fixed-length global feature vector, which serves as the fingerprint representation of the core skeleton. In this process, feature extraction and dimensionality transformation achieve a precise description of the molecular skeleton's structural boundaries. Based on this, a feature representation model of the molecule in chemical space is reconstructed, enabling the originally unstructured core skeleton to participate in subsequent clustering analysis in a numerical and computable form. Based on the constructed skeleton feature vector, the Tanimoto similarity between any two skeletons is further calculated. This similarity measures the proportion of common structural features in two fingerprints out of all appearing features, with a value between 0 and 1; a higher value indicates greater structural similarity between the two skeletons. The calculation formula is: , in, This represents the fingerprint generation function, used to convert the molecular skeleton into a binary bit string representation; and They represent the first The and the first The core framework of a molecule; This represents the intersection operation of two fingerprint bit strings, that is, counting the number of bits that are both 1; This represents the union operation of two fingerprint bit strings, that is, counting at least one bit that is 1. The numerator of this formula represents the number of structural features shared by the two skeletons, while the denominator represents the total number of all structural features that appear in the two skeletons. The ratio of the two ratios intuitively reflects the degree of structural similarity between the two skeletons.

[0041] After constructing a similarity matrix based on the similarity of all molecular pairs, the molecular set is divided into multiple skeleton clusters using an unsupervised clustering algorithm. Each cluster represents a chemical subspace with similar core structures.

[0042] Specifically, based on the similarity matrix Unsupervised clustering algorithms, such as agglomerative hierarchical clustering, are used to divide the molecular set into K skeleton clusters. Each cluster represents a chemical subspace with a similar core structure. This process transforms the originally disordered set of candidate molecules into a structured spatial representation, allowing information about the skeletal cluster to which a molecule belongs to to be explicitly used for subsequent diversity guidance. Through clustering, an ordered partitioning of the chemical space is achieved, thereby solving the problem of skeletal homogenization caused by the lack of explicit guidance on the molecular skeleton.

[0043] Furthermore, after completing the scaffold clustering, each candidate molecule is scored with multi-dimensional attributes, including three dimensions: bioactivity score, structural diversity score, and scaffold cluster novelty reward. Existing molecule generation methods typically focus only on a single bioactivity indicator or simply linearly combine multiple indicators, failing to fully consider the importance of molecular structural diversity and scaffold novelty in drug design. This often results in generation results limited to a certain type of advantageous scaffold, making it difficult to discover truly innovative lead compounds. In this embodiment of the invention, a comprehensive evaluation of three dimensions is introduced, providing a foundation for subsequent multi-objective optimization.

[0044] Bioactivity scores are used to measure the potential efficacy of molecules. Specifically, a pre-trained quantitative structure-activity relationship (QSPR) model is used to predict the activity value of a molecule against a target site. The QSPR model can be trained based on known activity data using architectures such as random forests, graph neural networks, or deep neural networks, and can make rapid and accurate activity predictions for newly generated molecules.

[0045] The predicted activity values ​​of all molecules in the current batch are subjected to min-max normalization, mapping the scores to the interval between 0 and 1. The normalization formula is as follows: , in, The predicted activity value of molecule M is given. and These are the minimum and maximum predicted activity values ​​for molecules in the current batch, respectively. This normalization process ensures comparability of scores between different batches, with higher scores indicating higher relative activity of the molecule within the current batch.

[0046] Structural diversity score measures the structural uniqueness of a molecule, preventing the model from repeatedly generating a large number of structurally similar molecules and thus ensuring the diversity of the generated results. Specifically, it calculates the average dissimilarity of a molecule compared to other molecules in the same batch, using the following formula: , in, For molecules and Tanimoto similarity of complete molecular fingerprints This represents the total number of molecules in the current batch.

[0047] The larger the value, the greater the structural difference between the molecule and other molecules in the batch, and the stronger its structural novelty. The introduction of diversity score can effectively suppress the tendency of the model to get stuck in local optima and repeatedly generate similar molecules during the optimization process.

[0048] The novelty incentive for molecular skeletons encourages models to explore sparsely populated molecular skeletons, thereby guiding the generative process to expand into a broader chemical space. In medicinal chemistry, the molecular skeleton determines the overall configuration and key physicochemical properties of a molecule; different skeleton types often correspond to different bioactive profiles.

[0049] In this embodiment of the invention, a novelty reward is calculated for each molecule based on the backbone cluster information obtained from clustering: Let the molecule Belongs to the skeleton cluster The novelty reward for the molecule's skeletal cluster is: , in, This represents the number of molecules in the current batch of this cluster. The preset adjustment coefficient is used. The formula reflects a reward mechanism that is inversely proportional to the cluster size. That is, molecules in a skeletal cluster with fewer molecules receive a higher novelty reward, thereby guiding the model to actively explore rare skeletal types that have not yet been fully explored, fundamentally solving the problem of skeletal homogenization.

[0050] Furthermore, after obtaining the scores for each molecule across its three dimensions, Pareto front analysis is performed to identify the Pareto optimal set of molecules. Pareto optimization is an important method in multi-objective decision-making, seeking a balance among multiple potentially conflicting objectives so that an improvement in one objective does not lead to a decline in another. In drug molecule design, bioactivity, structural diversity, and scaffold novelty are often interdependent. Simply pursuing high activity may result in molecules with homogeneous structures, while overemphasizing diversity may sacrifice activity. Pareto optimization can automatically identify the set of molecules that achieves optimal balance across all three dimensions.

[0051] Specifically, the bioactivity score, structural diversity score, and skeletal cluster novelty reward of each candidate molecule are constructed into a three-dimensional attribute feature vector, mapping each candidate molecule to a data point in the three-dimensional attribute space. Based on this, the dominance relationship between molecules is defined: for any two candidate molecules M1 and M2, if the score of M1 in all three dimensions is not lower than that of M2, and the score of M1 in at least one dimension is strictly higher than that of M2, then M1 is determined to dominate M2, and M2 is the dominated molecule.

[0052] All candidate molecules in the current batch are sorted by non-dominated order, and all non-dominated molecules that are not dominated by any other candidate molecules are screened out. This set of non-dominated molecules is identified as the first Pareto front, which is the Pareto optimal set of molecules. Molecules in the Pareto optimal set achieve the best balance in the current batch in terms of biological activity, structural diversity, and skeletal novelty.

[0053] Furthermore, after identifying the Pareto optimal set of molecules, the composite reward value for each candidate molecule is calculated. This composite reward value consists of two parts: the base weighted score and the frontier reward bonus. The calculation formula is as follows: , in, To score bioactivity, Structural diversity score As a reward for the novelty of the skeleton cluster, , , The preset weighting coefficients satisfy... This is used to balance the relative importance of the three dimensions in the basic reward. This is an indicator function, taking a value of 1 when the molecule belongs to the first Pareto front, and 0 otherwise. This represents the frontier reward intensity coefficient. In this composite reward calculation, the skeleton cluster novelty reward... The hierarchical reward coefficients of the cluster to which the molecule belongs (i.e., the intra-cluster reward adjustment coefficients corresponding to different levels) have been integrated, thereby enabling the hierarchical reward coefficients to pass through Indirectly adjust the final composite reward value to provide additional incentives for higher-level molecular clusters.

[0054] The compound reward calculation formula reflects two levels of consideration: the base weighted score reflects the molecule's comprehensive performance across the three dimensions; the frontier reward bonus provides additional positive incentives to Pareto-optimal molecules, enabling these molecules, which achieve optimal balance across multiple objectives, to receive higher compound reward values. Through this compound reward mechanism, the reinforcement learning optimization process can balance the comprehensive performance across the three dimensions of biological activity, structural diversity, and scaffold novelty, while also highlighting and reinforcing high-quality molecules that achieve optimal balance across multiple objectives. This guides the policy network to favor generating these high-quality drug candidate molecules in subsequent iterations.

[0055] Furthermore, after calculating the composite reward value for each candidate molecule, the parameters of the policy network are updated with this composite reward value as the optimization guide. The updated policy network is then used as the current policy network for the next iteration. This process is repeated until the policy network performance converges or the preset number of iterations is reached, thus obtaining the optimized policy network.

[0056] In the reinforcement learning framework, the molecular generation process is modeled as a sequence decision-making task. The policy network selects the next character as the action based on the current state of the generated character sequence, until a complete simplified linear input canonical sequence of molecules is generated.

[0057] Existing reinforcement learning-based molecule generation methods often use a single attribute as a reward or simply linearly weight multiple attributes, ignoring the interrelationships between attributes. This leads to the model easily favoring one attribute at the expense of others during optimization. For example, if only bioactivity is used as a reward, the model will quickly converge to a few highly active skeletons, generating a large number of structurally similar molecular variants, causing severe skeleton homogenization. If activity and diversity are simply weighted and summed, the weight coefficients are highly dependent on human experience and are difficult to adapt to the dynamic needs of different targets and optimization stages. Furthermore, traditional policy gradient methods such as the REINFORCE algorithm may suffer from training instability during updates. Excessively large update steps may cause policy collapse, resulting in the generated molecules suddenly becoming ineffective or falling into a single mode; too small a step size leads to slow convergence, requiring numerous iterations to obtain a better policy.

[0058] To overcome these problems, in this embodiment of the invention, the Proximal Policy Optimization (PPO) algorithm is used to optimize the policy network parameters. The update process is as follows. The near-end policy optimization algorithm is a trust region-based policy optimization method that limits the magnitude of policy changes in each update step to prevent a sharp decline in policy performance due to excessively large single updates. The near-end policy optimization algorithm achieves this goal through a pruning mechanism, with the objective function being: , in, The probability ratio indicates that the current policy network and the policy network before the update are in the same state. Select the same action below The probability ratio, For the current policy network to be optimized, The probability ratio represents the policy network before the update, reflecting the direction and magnitude of the policy update: when the probability ratio is greater than 1, it indicates that the current policy tends to choose the action more frequently, that is, the probability of the action has increased compared to before the update; when it is less than 1, it indicates that the tendency has decreased, that is, the probability of the action has decreased. This is an estimate of the advantage function, used to measure performance in state Select action The dominance function measures the degree of superiority or inferiority of an action relative to the average level. It is typically calculated using the generalized dominance estimation method, which considers the difference between the current reward and the expected future reward. A positive value indicates that the action is better than the average, while a negative value indicates that it is worse than the average. In the molecular generation scenario, the dominance function reflects the contribution of choosing a specific character to the final molecular composite reward given the current generated sequence. For the clipping function, These are the preset cropping parameters.

[0059] The objective function adaptively constrains the policy update magnitude by taking the minimum value between the pruned objective and the original objective. Specifically, when the dominance function is positive, the objective function is limited to the pruned value to prevent the policy from excessively increasing the probability of the action; when the dominance function is negative, the objective function is also limited to prevent the policy from excessively decreasing the probability of the action. This mechanism ensures that when the probability ratio exceeds the pruning range, the objective function no longer continues to grow with the increase or decrease of the probability ratio, thereby suppressing the magnitude of the policy update. In this way, the proximal policy optimization algorithm can carefully improve the policy in each update step, ensuring both the effective optimization direction of the policy and avoiding policy collapse due to excessively large single update magnitude, thus ensuring the stability of the training process.

[0060] In the specific implementation, each iteration's update process includes multiple optimization steps. The current policy network generates a batch of candidate molecules, typically numbering in the thousands. These molecules, after the evaluation module calculates the compound reward value, together with the states and action sequences generated during the generation process, constitute the training samples. Furthermore, the advantage function estimate is calculated based on these samples. and the probability ratio of the old and new strategies .

[0061] A proximal policy optimization objective function is constructed, and the policy network parameters are updated multiple times using gradient ascent, with gradients calculated using a mini-batch of samples during each update. During the update process, an entropy regularization term is typically added to encourage the policy to maintain a degree of exploratory behavior and prevent premature convergence to a deterministic policy. After multiple parameter updates, a new policy network is obtained.

[0062] The updated policy network is used as the current policy network for the next iteration. The closed-loop operation of generation, evaluation, and updating is repeated until the policy network performance converges or the preset number of iterations is reached. When the termination condition is met, the iteration stops, and the resulting policy network is the optimized policy network.

[0063] Through multiple rounds of iterative optimization, the optimized strategy network has learned to balance three dimensions—biological activity, structural diversity, and scaffold novelty—when generating molecules. Compared to the initial strategy network, the optimized network no longer merely replicates the molecular distribution of the training set but possesses targeted optimization capabilities based on target attributes. In each generation round, the optimized strategy network can consistently and stably generate drug candidate molecules of excellent overall quality. These molecules exhibit high potential in biological activity, structural diversity, and novel scaffold types, providing a high-quality candidate library for subsequent drug screening and lead compound discovery.

[0064] By employing an iterative optimization mechanism, skeletal clustering analysis and Pareto multi-objective evaluation are deeply integrated into the core decision-making loop of reinforcement learning. This achieves closed-loop optimization from molecular generation, skeletal clustering, multi-objective evaluation to policy update. In this embodiment, skeletal cluster information directly participates in the calculation of composite rewards, thereby influencing the update direction of the policy network and providing explicit guidance for the molecular skeletal structure. The model can proactively identify rare skeletal types and provide reward reinforcement during the optimization process, thus maintaining the ability to explore a broad chemical space while pursuing high biological activity, solving the problem of skeletal homogenization, and enhancing the breadth and depth of chemical space exploration.

[0065] The candidate molecules generated by the optimization strategy network are used as the final output drug candidate molecules. The optimization strategy network is then used to generate a batch of candidate molecules, the number of which can be set according to subsequent screening needs, such as generating thousands or tens of thousands of molecules. These generated molecules undergo grammatical checking and valence verification to form the final drug candidate molecule set. The molecules in this set possess good predictive biological activity, structural diversity, and scaffold novelty, providing a high-quality candidate library for subsequent drug screening, activity testing, and lead compound optimization.

[0066] Example 2: Based on the same inventive concept as in Embodiment 1, the present invention also provides a drug molecule generation system with enhanced backbone clustering, comprising the following modules: The data preprocessing and model pretraining module is used to preprocess the molecular dataset, generate a model using the pre-trained sequence of the preprocessed molecular dataset, and obtain the initial policy network. The iterative optimization module is used to take the initial policy network as the current policy network for the first round of iteration, generate candidate molecules from the current policy network, extract the core backbone of each candidate molecule, and cluster the candidate molecules based on backbone similarity to divide them into multiple backbone clusters. Based on the scaffold cluster, each candidate molecule is scored with multi-dimensional attributes, including bioactivity score, structural diversity score, and scaffold cluster novelty bonus. Pareto front analysis is performed based on the multi-dimensional scores of all candidate molecules to identify the Pareto optimal molecular set. The composite reward value is calculated based on the multi-dimensional scores of each candidate molecule and whether it belongs to the Pareto optimal molecular set. Guided by the composite reward value, the parameters of the policy network are updated, and the updated policy network is used as the current policy network for the next iteration until the performance of the policy network converges or the preset number of iterations is reached, thus obtaining the optimized policy network. The result output module is used to take the candidate molecules generated by the optimization strategy network as the final output drug candidate molecules.

[0067] The data preprocessing and model pretraining modules are used to preprocess the molecular dataset. Specifically, this includes removing molecules with invalid SMILES syntax or unreasonable valences, converting valid SMILES into standard canonical form to ensure representation uniqueness, and removing duplicate molecules to obtain a preprocessed training dataset. Then, the dataset is used to pretrain sequences to generate a model, resulting in an initial policy network. This policy network serves as the prior knowledge base for subsequent reinforcement learning processes.

[0068] The iterative optimization module is the core of the system, used to perform closed-loop iterative optimization of "generation-evaluation-update". In each iteration, the module uses the current policy network as a molecule generator, autoregressively sampling and generating a batch of candidate molecules. In the initial iteration, the current policy network is the initial policy network output by the data preprocessing and model pretraining modules; in subsequent iterations, it is the policy network updated in the previous round. Then, the module extracts the core backbone of each candidate molecule, retaining the ring systems and atoms and bonds on the shortest paths connecting the ring systems, removing all side chains and substituents, and converting the core backbone into a molecular fingerprint. Based on the Tanimoto similarity between fingerprints, a similarity matrix is ​​constructed, and an unsupervised clustering algorithm is used to divide the candidate molecules into multiple backbone clusters, each representing a chemical subspace with similar core structures. Then, based on the Pareto front analysis results (i.e., the non-dominated level of multi-objective attribute scores) of the candidate molecules within each cluster, the molecular clusters are hierarchically ranked: clusters with higher Pareto priority are placed at the top of the hierarchy, and clusters with lower priority are placed at the bottom, thus constructing a hierarchical molecular cluster structure with different priorities. Different levels correspond to different intra-cluster reward adjustment coefficients, with higher-level molecular clusters enjoying higher reward coefficients. Based on this structure, each candidate molecule is scored with multi-dimensional attributes, and the optimal molecular set is identified based on Pareto front analysis. The composite reward value is then calculated and used for parameter updates in the policy network.

[0069] Based on the partitioned skeletal clusters, the iterative optimization module performs hierarchical ranking of each molecular cluster according to the Pareto front analysis results (i.e., the non-dominated level of multi-objective attribute scores) of candidate molecules within each cluster: molecular clusters with higher Pareto priority are placed at the upper level of the hierarchical structure, and clusters with lower priority are placed at the lower level, thus constructing a hierarchical molecular cluster structure with different priorities. Different levels correspond to different intra-cluster reward adjustment coefficients, with molecular clusters at higher levels enjoying higher reward coefficients. Based on this hierarchical molecular cluster structure, the iterative optimization module performs multi-dimensional attribute scoring on each candidate molecule. The multi-dimensional attribute scoring includes bioactivity score, structural diversity score, and skeletal cluster novelty reward. The bioactivity score is obtained by predicting the molecular activity value through a pre-trained quantitative structure-activity relationship model and performing batch-level normalization. The structural diversity score is obtained by calculating the average dissimilarity of the molecule to other molecules in the same batch. The novelty reward of the backbone cluster is calculated inversely proportional to the current size of the backbone cluster to which the molecule belongs, and further weighted by the hierarchical position of the cluster to which the molecule belongs. Molecules in higher-level clusters receive additional novelty reward bonuses. Then, the module maps the three-dimensional scores of all candidate molecules to points in the three-dimensional attribute space, and uses a fast non-dominated sorting algorithm to perform Pareto front analysis to identify the first Pareto front that is not dominated by any other molecule, which is taken as the Pareto optimal molecule set. Then, based on the multi-dimensional scores of each candidate molecule and whether it belongs to the optimal molecule set, a composite reward value is calculated. This composite reward value consists of the basic weighted score and the front reward bonus, so that molecules that achieve optimal balance in multiple objectives receive higher rewards.

[0070] The iterative optimization module uses the calculated composite reward value as the optimization guide. It employs a reinforcement learning algorithm (preferential proximal policy optimization) to update the parameters of the policy network. A pruning mechanism is used to limit the policy update magnitude, maximizing the expected reward while ensuring training stability. The updated policy network serves as the current policy network for the next iteration. The above generation, clustering, evaluation, and update process is repeated until the policy network performance converges or the preset number of iterations is reached, at which point the optimized policy network is obtained.

[0071] The output module takes the candidate molecules generated by the optimized strategy network as the final output drug candidates. These molecules exhibit excellent performance in terms of biological activity, structural diversity, and scaffold novelty, providing a high-quality candidate library for drug screening and lead compound discovery. Through the modular design described above, this system deeply integrates scaffold clustering analysis and Pareto multi-objective evaluation into the core decision loop of reinforcement learning. The modules have low coupling, clear interfaces, and good versatility and scalability.

[0072] Example 3: The present invention also provides a computer storage medium storing a computer software product, the computer software product including a plurality of instructions for causing a computer device to execute the enhanced backbone clustering drug molecule generation method described in Embodiment 1.

[0073] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for generating drug molecules using enhanced backbone clustering, characterized in that, include: The molecular dataset is preprocessed, and the pre-trained sequences of the preprocessed molecular dataset are used to generate a model, resulting in an initial policy network. The initial policy network is used as the current policy network for the first iteration, and candidate molecules are generated from the current policy network. The core backbone of each candidate molecule is extracted, and the candidate molecules are clustered based on backbone similarity to form multiple backbone clusters; Based on the scaffold cluster, each candidate molecule is scored with multi-dimensional attributes, including bioactivity score, structural diversity score, and scaffold cluster novelty bonus. Pareto front analysis is performed based on the multi-dimensional scores of all candidate molecules to identify the Pareto optimal molecular set. The composite reward value is calculated based on the multi-dimensional scores of each candidate molecule and whether it belongs to the Pareto optimal molecular set. Guided by the composite reward value, the parameters of the policy network are updated, and the updated policy network is used as the current policy network for the next iteration until the performance of the policy network converges or the preset number of iterations is reached, thus obtaining the optimized policy network. The candidate molecules generated by the optimization strategy network are used as the final output drug candidate molecules.

2. The method for generating drug molecules using enhanced backbone clustering according to claim 1, characterized in that: The method for obtaining the initial policy network by pre-training the sequence generation model using the preprocessed molecular dataset is as follows: A neural network based on a self-attention mechanism is constructed as the sequence generation model. The simplified linear input canonical sequences from the preprocessed molecular dataset are used as the training dataset. Maximum likelihood estimation is employed as the pre-training objective, maximizing the conditional log-likelihood of the sequence. The mathematical expression is: ， in, This represents a simplified linear input canonical sequence of a molecule, where T is the sequence length. For the t-th character, This represents the historical sequence preceding the t-th character. For policy networks, For model parameters, The training dataset is used; the model parameters are iteratively updated, and after training, the learned model parameters are used as the initial policy network.

3. The method for generating drug molecules using enhanced backbone clustering according to claim 1, characterized in that: The method for extracting the core skeleton of each candidate molecule is as follows: the simplified linear input canonical representation of each molecule is analyzed to identify all ring systems in the molecule, retain the ring systems and the atoms and bonds on the shortest path connecting the ring systems, remove all non-cyclic side chain atoms and substituents, and obtain a simplified structure containing only ring systems and linkers, which serves as the core skeleton of the molecule.

4. The method for generating drug molecules using enhanced backbone clustering according to claim 1, characterized in that: The method for clustering the candidate molecules into multiple backbone clusters based on backbone similarity is as follows: The core backbone of each molecule is converted into a molecular fingerprint, and the similarity between any two backbones is calculated. and The formula for the Tanimoto similarity between them is: ， in, This represents the fingerprint generation function. and Let represent the intersection and union of the sets respectively; construct a similarity matrix based on the similarity of all molecular pairs, and divide the molecular set into multiple skeleton clusters through unsupervised clustering, with each cluster representing a chemical subspace with similar core structures.

5. The method for generating drug molecules using enhanced backbone clustering according to claim 1, characterized in that: Based on the scaffold cluster, each candidate molecule is scored with multi-dimensional attributes, including bioactivity score, structural diversity score and scaffold cluster novelty reward. Specifically, the scaffold cluster is hierarchically sorted according to the Pareto non-dominated level of the candidate molecules in each scaffold cluster, and hierarchical molecular cluster structures with different priorities are constructed. Different levels correspond to different intra-cluster reward regulation coefficients. Based on the hierarchical molecular cluster structure, each candidate molecule is scored with multi-dimensional attributes, including bioactivity score, structural diversity score, and backbone cluster novelty reward. The backbone cluster novelty reward is dynamically calculated based on the hierarchical position of the cluster to which the molecule belongs and the sparsity within the cluster.

6. The method for generating drug molecules using enhanced backbone clustering according to claim 1, characterized in that: Pareto front analysis based on the multi-dimensional scores of all candidate molecules is used to identify the Pareto optimal molecular set. The method is as follows: A three-dimensional attribute feature vector is constructed from the bioactivity score, structural diversity score, and skeletal cluster novelty reward of each candidate molecule, mapping each candidate molecule to a data point in the three-dimensional attribute space. The dominance relationship between molecules is defined: for any two candidate molecules M1 and M2, if M1's scores in all three dimensions (bioactivity score, structural diversity score, and skeletal cluster novelty reward) are not lower than M2's, and M1's score in at least one dimension is strictly higher than M2's, then M1 is determined to dominate M2, and M2 is the dominated molecule. All candidate molecules in the current batch are sorted for non-dominated molecules, and all non-dominated molecules that are not dominated by any other candidate molecules are selected. The set of non-dominated molecules is identified as the first Pareto front, which is the Pareto optimal molecular set.

7. The method for generating drug molecules using enhanced backbone clustering according to claim 1, characterized in that: The method for calculating the composite reward value for each candidate molecule M is as follows: ， in, To score bioactivity, To score structural diversity, As a reward for the novelty of the skeleton cluster, , , The preset weighting coefficients satisfy... ; This is an indicator function; its value is 1 when the molecule is at the Pareto front, and 0 otherwise. It is the Pareto optimal molecular set; This refers to the frontier reward intensity coefficient.

8. The method for generating drug molecules using enhanced backbone clustering according to claim 1, characterized in that: The method for updating the parameters of the policy network, guided by the aforementioned composite reward value, is as follows: Reinforcement learning algorithm is used to update the strategy network parameters. The objective function is: ， in, For probability ratios, This is the estimated value of the dominance function. For the clipping function, These are the trimming parameters; By maximizing this objective function, the policy network is guided to generate drug candidate molecules with high composite reward values ​​while ensuring training stability.

9. The method for generating drug molecules using enhanced backbone clustering according to claim 1, characterized in that: The method for preprocessing the molecular dataset is as follows: perform syntax checking and valence validity verification on the simplified linear input canonical representation of each molecule, and remove invalid molecules; convert the valid simplified linear input canonical representation of the molecules into the standard canonical form to ensure uniqueness; remove duplicate molecules to obtain the preprocessed training dataset.

10. A drug molecule generation system with enhanced backbone clustering, characterized in that, Includes the following modules: The data preprocessing and model pretraining module is used to preprocess the molecular dataset, generate a model using the pre-trained sequence of the preprocessed molecular dataset, and obtain the initial policy network. The iterative optimization module is used to take the initial policy network as the current policy network for the first round of iteration and generate candidate molecules from the current policy network. The core backbone of each candidate molecule is extracted, and the candidate molecules are clustered based on backbone similarity to form multiple backbone clusters; Based on the scaffold cluster, each candidate molecule is scored with multi-dimensional attributes, including bioactivity score, structural diversity score, and scaffold cluster novelty bonus. Pareto front analysis is performed based on the multi-dimensional scores of all candidate molecules to identify the Pareto optimal molecular set. The composite reward value is calculated based on the multi-dimensional scores of each candidate molecule and whether it belongs to the Pareto optimal molecular set. Guided by the composite reward value, the parameters of the policy network are updated, and the updated policy network is used as the current policy network for the next iteration until the performance of the policy network converges or the preset number of iterations is reached, thus obtaining the optimized policy network. The result output module is used to take the candidate molecules generated by the optimization strategy network as the final output drug candidate molecules.

Images