A plug-and-play lead compound optimization method and system

Abstract

The application discloses a plug-and-play lead compound optimization method and system, and belongs to the technical field of computer-aided drug design. The method comprises the following steps: obtaining target protein pocket information and known molecular fragment information, fixing atomic coordinates and atomic types as constraint conditions, initializing an unknown molecular region, and generating an initial molecular structure; inputting the target protein pocket information and the initial molecular structure into a model, obtaining a first sampling direction and a second sampling direction; iteratively updating atomic coordinates according to the first sampling direction and the constraint conditions through a continuous mode sampler, and iteratively updating atomic types according to the second sampling direction and the constraint conditions through a discrete mode sampler; and when the iterative update meets a preset termination condition, outputting a complete molecular structure containing the known molecular fragment and capable of being combined with the target protein pocket. The application considers both generation efficiency and molecular rationality without resampling, and significantly improves lead compound optimization efficiency and flexibility.

Description

Technical Field

[0001] This invention relates to the field of computer-aided drug design technology, and more specifically to a plug-and-play method and system for optimizing lead compounds. Background Technology

[0002] Completing known molecular fragments into complete drug molecules capable of effectively binding to specific protein pockets is a crucial step in drug discovery. Typically, this process begins with a lead compound, and through systematic structural modification and completion, enhances the overall binding affinity, selectivity, and physicochemical properties between the compound and the target protein. Based on the specific sites and strategies of modification and completion, this task can be subdivided into four typical subtasks: linker design, fragment growth, scaffold transition, and side-chain modification.

[0003] Traditional lead compound optimization methods primarily rely on chemists' experience-based rational design and high-throughput screening experiments. However, these methods are often costly, time-consuming, and difficult to conduct systematic and comprehensive exploration in high-dimensional molecular spaces. In recent years, with the development of deep learning technology, generative models have gradually demonstrated powerful potential, capable of completing multiple molecular optimization tasks within a unified framework, significantly improving the speed and innovation of drug design. Currently, the main technical routes for lead compound optimization using deep learning technology can be broadly divided into two categories.

[0004] The first type of method involves training the generative model from scratch on different sub-tasks. By constructing specific training datasets or introducing targeted masking strategies, the model can learn different types of structural modification patterns and optimization strategies. This type of method typically revolves around the four typical sub-tasks mentioned above. It requires preparing separate training data or masking strategies for each sub-task, resulting in a complex overall training process, high computational resource consumption, and long model optimization cycles.

[0005] The second type of method employs a generative framework based on Structure-Based Drug Design (SBDD), utilizing inpainting techniques to locally complete and optimize existing molecular structures. To ensure the effectiveness and diversity of generated samples, the model often requires multiple resampling operations or energy-constrained screening during the inference phase to guarantee the quality and rationality of the generated molecules. This significantly increases the computational cost during the inference phase, making it difficult to balance efficiency and generation performance. Summary of the Invention

[0006] To address the issues of low efficiency and high computational resource consumption in lead compound optimization during drug design, this invention proposes a plug-and-play lead compound optimization method and system. Based on a generative framework for structure-based drug discovery, lead compound optimization is achieved. The sampler constructs variational distributions for atomic coordinates and atomic types, directly deriving the gradient closed-form solution of the divergence with respect to the variational distribution parameters. This eliminates the need for resampling during the inference process, significantly reducing computational overhead while ensuring the quality and chemical rationality of the generated molecules, thus balancing efficiency and generative performance.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a plug-and-play lead compound optimization method, comprising the following steps: S1. Obtain target protein pocket information and known molecular fragment information, wherein the known molecular fragment information includes determined atomic coordinates and atomic types; fix the atomic coordinates and atomic types as constraints, and initialize the atomic coordinates and atomic types in the unknown molecular region to generate the initial molecular structure to be sampled; S2. Input the target protein pocket information and initial molecular structure into the pre-trained generation model to obtain the first sampling direction of the pre-trained generation model for continuous modal output and the second sampling direction for discrete modal output; S3. Based on the first sampling direction and the constraint conditions, the atomic coordinates of the molecule to be sampled are iteratively updated using a continuous modal sampler so that the atomic coordinates approximate the conditional distribution that satisfies the constraint conditions. S4. Based on the second sampling direction and the constraint conditions, the atom types of the molecule to be sampled are iteratively updated using a discrete modal sampler so that the atom types approximate the conditional distribution that satisfies the constraint conditions. S5. When the iterative updates of the continuous modal sampler and the discrete modal sampler both meet the preset termination condition, the complete molecular structure containing the known molecular fragment and capable of binding to the target protein pocket is output.

[0009] Furthermore, in step S1, fixing the atomic coordinates and atomic type as constraints specifically includes: The atomic coordinates and atom types of each atom contained in the known molecular fragment remain unchanged during the iterative update process, and the known molecular fragment serves as an immutable structural anchor point in the complete molecular structure.

[0010] Furthermore, in step S1, the atomic coordinates and atom types in the unknown molecular region are initialized, specifically including: The unknown molecular region is the region containing atoms other than those contained in the known molecular fragment in the complete molecular structure; The atomic coordinates of each atom in the unknown molecular region are randomly initialized using a Gaussian distribution; The atom types of each atom in the unknown molecular region are randomly initialized using a uniform distribution.

[0011] Furthermore, in step S2, the pre-trained generative model is a diffusion generative model trained under the framework of structure drug discovery. The model takes the three-dimensional structure of the protein pocket as a conditional input and outputs the three-dimensional molecular structure and atom type that can bind to the protein pocket. The first sampling direction is the noise prediction direction of the pre-trained generation model for the coordinates of each atom in the molecule to be sampled, and the second sampling direction is the probability prediction distribution of each atom type in the molecule to be sampled by the pre-trained generation model.

[0012] Furthermore, in step S3, the atomic coordinates of the molecule to be sampled are iteratively updated using a continuous modal sampler, specifically including: A continuous variational distribution is introduced for the atomic coordinates, which is used to approximate the true conditional distribution of the atomic coordinates under the constraints. Based on the first sampling direction output by the pre-trained generative model and the atomic coordinate constraints of the known molecular fragment, calculate the gradient of the divergence between the variational distribution and the true conditional distribution with respect to the variational distribution parameters; The parameters of the variational distribution are iteratively updated according to the gradient until a preset number of iterations is reached, so as to gradually correct the atomic coordinates.

[0013] Furthermore, the gradient formula for calculating the divergence between the variational distribution and the true conditional distribution with respect to the variational distribution parameters is as follows:

[0014] in, This represents a pre-trained neural network. Represents the atomic coordinates in a complete molecule. This represents the atomic coordinates in a known molecular fragment. Represents a continuous diffusion model The latent variable corresponding to time. Represents distribution The parameters, The hyperparameters representing the control constraint weights The mapping matrix represents the generated molecules. Atoms and fragment structures contained The mapping relationship of the atoms contained.

[0015] Furthermore, in step S4, the atom types of the molecule to be sampled are iteratively updated using a discrete modal sampler, specifically including: A discrete variational distribution is introduced for the atom type, which is used to approximate the true conditional distribution of the atom type under the constraints. Based on the second sampling direction output by the pre-trained generative model and the atomic type constraints of the known molecular fragment, calculate the gradient of the divergence between the variational distribution and the true conditional distribution with respect to the variational distribution parameters; The parameters of the variational distribution are iteratively updated according to the gradient until a preset number of iterations is reached, so as to achieve gradual correction of the atom type.

[0016] Furthermore, the gradient formula for calculating the divergence between the variational distribution and the true conditional distribution with respect to the variational distribution parameters is as follows:

[0017] in, Indicates the types of atoms in a complete molecule. Indicates the types of atoms in a known molecular fragment. Indicates the first The probability parameters of each category, Representing the discrete diffusion model The latent variable corresponding to time. This represents the latent variable at time 0 in the discrete diffusion model, i.e., the clean sample without noise. This means that all element values ​​are 0. The vector, This represents the forward process of the diffusion model. This represents the reverse process of the unconditional diffusion model. This represents the inverse process of the diffusion model corresponding to the variational distribution. This represents the marginal distribution of the unconditional diffusion model. Indicates the variational distribution Exported implicit variables The marginal distribution that is satisfied.

[0018] Furthermore, in step S5, the preset termination condition is that the number of iterations reaches a preset number of steps. The preset number of steps is the same as the total number of denoising steps used by the pre-trained generative model during the training phase, and the learning rate for each iteration is set to 0.1.

[0019] In a second aspect, embodiments of the present invention provide a plug-and-play lead compound optimization system, which applies the method described in the first aspect, the system comprising: An initialization module is used to acquire target protein pocket information and known molecular fragment information, wherein the known molecular fragment information includes determined atomic coordinates and atomic types; the atomic coordinates and atomic types are fixed as constraints, and the atomic coordinates and atomic types in the unknown molecular region are initialized to generate the initial molecular structure to be sampled; The pre-trained generative model module is used to input the target protein pocket information and initial molecular structure into the pre-trained generative model, and obtain the first sampling direction of the pre-trained generative model for continuous modal output and the second sampling direction for discrete modal output; A continuous modal sampler is used to iteratively update the atomic coordinates of the molecule to be sampled based on the first sampling direction and the constraint conditions, so that the atomic coordinates approximate the conditional distribution that satisfies the constraint conditions. A discrete modal sampler is used to iteratively update the atom type of the molecule to be sampled based on the second sampling direction and the constraint conditions, so that the atom type approximates the conditional distribution that satisfies the constraint conditions. The output module is used to output a complete molecular structure containing the known molecular fragment and capable of binding to the target protein pocket when the iterative updates of the continuous modal sampler and the discrete modal sampler both meet the preset termination condition.

[0020] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: The sampler proposed in this invention can be directly mounted onto any pre-trained diffusion generation model based on structure-based drug discovery without retraining or fine-tuning, enabling task transfer from unconditional molecule generation to conditional lead compound optimization. This plug-and-play capability significantly improves the model's applicability and scalability across different task scenarios.

[0021] This invention constructs variational distributions for atomic coordinates and atomic types respectively, and directly derives the closed-form solution of the gradient of the KL divergence with respect to the variational distribution parameters. This allows the sampler to simultaneously consider global generation trends and local structural constraints in a single forward propagation. The inference process does not require resampling, significantly reducing computational overhead while ensuring the quality and chemical rationality of the generated molecules.

[0022] This invention, within a unified mathematical framework, enables the simultaneous completion of four lead compound optimization subtasks—connector design, fragment growth, backbone transition, and side chain modification—through the collaborative iteration of continuous and discrete modal samplers. It eliminates the need for separate model architectures, training strategies, or data masking schemes for different subtasks, significantly simplifying model deployment and application processes and providing a highly efficient, universal, and interpretable automated tool for drug development. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 This is a flowchart of the plug-and-play lead compound optimization method provided in this embodiment of the invention.

[0025] Figure 2 This is a schematic diagram of the sampling stage of the plug-and-play lead compound optimization method provided in this embodiment of the invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1: like Figure 1 As shown, this embodiment of the invention discloses a plug-and-play lead compound optimization method, comprising the following steps: S1. Obtain target protein pocket information and known molecular fragment information, wherein the known molecular fragment information includes the determined atomic coordinates and atomic types; fix the atomic coordinates and atomic types as constraints, and initialize the atomic coordinates and atomic types in the unknown molecular region to generate the initial molecular structure to be sampled; fix the known molecular fragment information as an immutable structural anchor point to ensure that the generated molecule always retains the key active structure and avoids the shift or loss of key pharmacophore groups during the optimization process.

[0028] S2. Input the target protein pocket information and initial molecular structure into the pre-trained generation model to obtain the first sampling direction for the continuous modal output and the second sampling direction for the discrete modal output of the pre-trained generation model; use the pre-trained generation model directly as a conditional information extractor, without retraining or fine-tuning, to obtain the bimodal guidance direction for the current constraints, achieving true plug-and-play functionality.

[0029] S3. Based on the first sampling direction and the constraint conditions, the atomic coordinates of the molecule to be sampled are iteratively updated using a continuous modal sampler so that the atomic coordinates approximate the conditional distribution that satisfies the constraint conditions. The continuous modal sampler directly calculates the gradient closed-form solution of the variational distribution, taking into account both global conformation rationality and local position constraints in a single forward propagation. Spatial structure optimization can be completed without multiple resampling, greatly reducing computational overhead.

[0030] S4. Based on the second sampling direction and the constraint conditions, the atomic type of the molecule to be sampled is iteratively updated by a discrete modal sampler so that the atomic type approximates the conditional distribution that satisfies the constraint conditions. The discrete modal sampler directly derives the gradient closed solution in the category distribution parameter space without relaxing the discrete variables into continuous variables, avoiding the generation of fidelity loss and ensuring the chemical rationality of the optimized molecule and the consistency of the fragment type.

[0031] S5. When the iterative updates of the continuous modal sampler and the discrete modal sampler both meet the preset termination conditions, a complete molecular structure containing the known molecular fragment and capable of binding to the target protein pocket is output. The iteration termination is controlled by a preset number of steps and learning rate to achieve efficient automatic convergence. The output complete molecule outperforms existing generative frameworks optimized for lead compounds in terms of comprehensive performance indicators such as binding affinity and drug-likeness.

[0032] To optimize lead compounds while balancing computational efficiency and generation performance, this invention first formalizes the problem as a conditional generation task. Given both the unconditional generation model and the observation model, a conditional distribution is constructed to efficiently sample target molecules, thereby improving generation speed while ensuring diversity and rationality. Subsequently, considering the discrete data modal characteristics of chemical molecules, an efficient discrete sampler is proposed, which significantly reduces computational overhead while maintaining sampling quality. Finally, this discrete sampler is combined with an existing continuous modal sampler to construct a unified hybrid sampling framework. Based on this framework, within the generative model system of structure-based drug discovery, it is possible to complete and modify existing molecular fragments, balancing computational efficiency and molecular design quality, and significantly improving the model's applicability and scalability in practical drug development scenarios.

[0033] The plug-and-play lead compound optimization method of the present invention will be described in detail below: Specifically, such as Figure 2As shown, in the sampling phase, the sampled molecules first need to be initialized. For known parts, their atomic coordinates and atom types are set to fixed given values ​​to ensure that the structural constraints of known fragments are preserved. For unknown molecular regions, Gaussian distributions are used to initialize the atomic coordinates, and uniform distributions are used to initialize the atom types. Subsequently, the sampled molecules undergo multi-step iterative updates guided by the sampler. In each step, the sampler comprehensively utilizes fragment structural constraint information and an unconditional generation model, i.e., a generation framework based on structure-based drug discovery, to estimate the likelihood gradient of the conditional distribution and adjust the generation direction of the sampled molecules accordingly. Through this gradient-guided iterative process, the model can gradually approximate the conditional distribution, thereby generating molecular structures that conform to structural constraints and chemical rationality.

[0034] The following section provides a mathematical model for the condition generation task corresponding to the above sampling process, given protein pocket information. Information related to molecular fragments Generate complete molecules containing a given fragment that can bind to the pocket. ,in Indicates the first The coordinates of each atom, Indicates the first Atom type of each atom These represent the number of atoms contained in the protein pocket, the molecular fragment, and the resulting complete molecule, respectively. This process can be viewed as a distribution... Sampling involves determining the probability distribution that a complete molecule should satisfy, given the known protein pockets and molecular fragments. A structure-based drug discovery framework can be viewed as a way to... Sampling is performed, which involves modeling the distribution satisfied by intact molecules that can bind to a specific protein pocket. Simultaneously, the observed distribution is independent of the protein pocket. ,Right now Therefore, the above problem can be described as having a known distribution. and How to achieve distribution Sampling.

[0035] Since atomic coordinates and atomic types belong to different modes, the discrete and continuous modes are first separated. The generation of atomic coordinates is considered as a distribution... The sampling, the generation of atom types is regarded as a distribution Sampling.

[0036] in, and These represent the coordinates of atoms contained in a complete molecule and a molecular fragment, respectively. and These represent the types of atoms contained in a complete molecule and a molecular fragment, respectively.

[0037] The unconditional distribution and observed distribution of the two are written as: and , and For unconditional distributions and A common approach is to use a diffusion model for modeling, i.e., a pre-trained neural network. and Fitting and The loss function of the pre-trained model is:

[0038]

[0039] in, and These represent the continuous discrete modal diffusion model. The latent variables corresponding to each time point can be obtained through a predefined noisy distribution. and For clean data Obtained by applying a perturbation.

[0040] Regarding the architecture of the pre-trained model, the plug-and-play sampler proposed in this invention has no additional requirements. It only needs to be trained on a small molecule and protein complex dataset, and the pre-trained model can generate small molecules that can bind to specific protein pockets.

[0041] The following is a derivation of the discrete mode sampler: The target distribution of the sample is Its parameters are unknown. A variational distribution is introduced. :

[0042] in, Represents category distribution, Represent the probabilities of each class in vector form. Since the observed distribution and the unconditional distribution are known, the variational distribution holds as follows:

[0043] Among them, the last item For the variational distribution, this is a constant term, therefore:

[0044] The objective is to solve for the variational distribution. ,right Sampling is performed to achieve the target distribution The purpose of sampling. Among them, unconditional distribution. The observed distribution is obtained by modeling from a pre-trained model. Its physical meaning is the type of atoms in a given complete molecule. The distribution of the types of atoms contained in a molecular fragment can be modeled using the following distribution form:

[0045] in, The stringency of the control over atom type constraints, This represents the mapping matrix between molecular fragments and complete molecules. A value of 1 indicates that an atom in the complete molecule belongs to a given molecular fragment, and a value of 0 indicates that atom in the complete molecule belongs to a given molecular fragment. This represents a constant vector, where all elements are constants. , This represents the total number of atom types. An optimization-based approach is used, which first initializes the variational distribution parameters. Solving for the gradient right By iterating, the variational distribution can be solved. .

[0046] The following theorem can be derived: Let the variational distribution be... Unconditional distribution Modeled by a pre-trained diffusion model, latent variables are introduced into the variational distribution. Its distribution is caused by noise disturbance. ( | This is obtained by adding noise, which is consistent with the noise addition process in the diffusion model. ( | The objective function is... For variational distribution parameters The gradient is:

[0047] in, Indicates the types of atoms in a complete molecule. Indicates the types of atoms in a known molecular fragment. Indicates the first The probability parameters of each category, Representing the discrete diffusion model The latent variable corresponding to time. This represents the latent variable at time 0 in the discrete diffusion model, i.e., the clean sample without noise. This means that all element values ​​are 0. The vector, This represents the forward process of the diffusion model. This represents the reverse process of the unconditional diffusion model. This represents the inverse process of the diffusion model corresponding to the variational distribution. This represents the marginal distribution of the unconditional diffusion model. Indicates the variational distribution Exported implicit variables The marginal distribution that is satisfied. For This distribution is known, and it defines the forward process of the diffusion model. It is the fitting target of a pre-trained neural network (unconditional generative model), which can be obtained from the output of the neural network. Distribution Bayes' theorem can be used Obtained. For It can be obtained through the following process.

[0048] Due to the noise addition process ( | The design of ) regardless What distribution does it follow? Latent variables It must obey a uniform distribution, therefore Then there is

[0049] for The following techniques can be used for calculation:

[0050] in, and All are known. This represents the output of the pre-trained neural network. Because... Given (although the minimization condition cannot be met), therefore ,thus Given. For It is uniformly distributed, and the distribution is known.

[0051] Similar to discrete mode samplers, variational distributions are also introduced for continuous mode samplers. The distribution is set to Gaussian, and the variance can be set to approach 0. The continuous modal sampler solves for the gradient. right By iterating, the variational distribution can be solved. .

[0052] For continuous modes, the following theorem holds: Set observation distribution Introducing latent variables into the variational distribution Its distribution is caused by noise disturbance. ( | The distribution obtained by adding noise is consistent with the noise addition process in the diffusion model, i.e. Then the objective function For variational distribution parameters The gradient is:

[0053] in, This represents a pre-trained neural network. Represents the atomic coordinates in a complete molecule. This represents the atomic coordinates in a known molecular fragment. Represents a continuous diffusion model The latent variable corresponding to time. Represents distribution The parameters, The hyperparameters representing the control constraint weights The mapping matrix represents the generated molecules. Atoms and fragment structures contained The mapping relationship of the atoms contained.

[0054] Based on the two theorems mentioned above, a continuous and discrete modal sampler is proposed. The sampler uses the derivation of the gradient of the distribution parameters using the KL divergence from the theorems, and obtains the corresponding gradient direction based on the output of the unconditionally generated model and the parameters of the current variational distribution. The variational distribution parameters are then updated along the gradient direction, and the optimal variational distribution is obtained after multiple iterations. Typically, the number of iterations is consistent with the number of denoising steps in the pre-trained diffusion model, and the learning rate is set to 0.1. Optimization stops after reaching the specified number of iterations.

[0055] This invention formalizes the lead compound optimization process as a conditional generation problem, and establishes a unified mathematical framework for the sampling process of discrete and continuous modes. For different types of data characteristics in the molecule, suitable variational distributions are introduced under the diffusion model framework, and corresponding latent variable representations are constructed for each mode, enabling the model to flexibly capture the dependency relationship between molecular structure and protein pockets. This invention directly estimates the gradient of the KL divergence with respect to the variational distribution parameters, avoiding resampling or numerical integration operations in traditional methods, greatly reducing the computational overhead in the sampling process. Based on this gradient expression, this invention iterates the variational distribution parameters and samples the optimal variational distribution to obtain a complete molecule with a specified substructure that can bind to the protein pocket, exhibiting good theoretical interpretability and ensuring the desirable properties of the designed molecule.

[0056] The following dataset verification further demonstrates the plug-and-play lead compound optimization method of this invention: This invention evaluates the proposed method using the publicly available dataset CrossDocked2020 and conducts comparative experiments with existing representative methods on four typical subtasks. The comparison methods include DELETE and MODA, while the sampler proposed in this invention is tested on three different pre-trained generative models: DiffSBDD, TargetDiff, and DiffBP.

[0057] In the experiment, the performance of the generated molecules was comprehensively evaluated from two aspects: First, it combines capability indicators, including VinaScore (Vina energy calculation results), VinaMin (Vina optimization results), and VinaDock (Vina docking results). Second, physicochemical properties, including drug-likeness (QED), lipid solubility (LogP), syntheticability (SA), and the degree to which it meets the "five principles of drug-likeness" (LPSK).

[0058] In addition, for Vina Dock, the relative improvement rate (MPBG) was calculated based on the docking results of the reference molecules in the dataset, and the ligand binding efficiency (LBE) was calculated based on the number of atoms in the generated molecule as supplementary evaluation indicators.

[0059] In the results presentation, bold indicates that the indicator achieved the best result, and underline indicates that the indicator was the second best. The results of each method for affinity and physicochemical properties under different task and model combinations are shown in Tables 1 and 2 below: Table 1: Results of Affinity-Related Indicators

[0060] Table 2: Results of relevant physicochemical properties

[0061] Ours indicates that the plug-and-play sampler proposed in this invention has been introduced into the corresponding pre-trained model. Delete and MODA are generative models designed directly for lead compound optimization tasks.

[0062] As shown in Table 1, in the four sub-tasks, the sampler method significantly outperformed Delete and MODA in affinity-related indices (VinaScore, Vina Min, Vina Dock), indicating that the sampler can effectively improve the binding ability between the generated molecule and the target pocket. Table 2 shows that the sampler method slightly outperformed the comparative models in QED (drug-likeness) and LPSK (degree of conformity to the five principles of drug-likeness), indicating that its generated molecules have a better balance in terms of overall drug feasibility; in terms of LogP (lipophilicity), the results of all methods are within a reasonable range; while in terms of SA (syntheticability), the Delete model performs relatively better.

[0063] Taking all evaluation metrics into account, the sampler proposed in this study can not only efficiently transfer generative models based on structure drug discovery to lead compound optimization tasks without retraining or fine-tuning, but also produces and modifies molecules that outperform current generative frameworks designed specifically for lead compound optimization in terms of overall performance.

[0064] Example 2: Based on the same inventive concept, embodiments of the present invention also provide a plug-and-play lead compound optimization system, the system comprising: An initialization module is used to acquire target protein pocket information and known molecular fragment information, wherein the known molecular fragment information includes determined atomic coordinates and atomic types; the atomic coordinates and atomic types are fixed as constraints, and the atomic coordinates and atomic types in the unknown molecular region are initialized to generate the initial molecular structure to be sampled; The pre-trained generative model module is used to input the target protein pocket information and initial molecular structure into the pre-trained generative model, and obtain the first sampling direction of the pre-trained generative model for continuous modal output and the second sampling direction for discrete modal output; A continuous modal sampler is used to iteratively update the atomic coordinates of the molecule to be sampled based on the first sampling direction and the constraint conditions, so that the atomic coordinates approximate the conditional distribution that satisfies the constraint conditions. A discrete modal sampler is used to iteratively update the atom type of the molecule to be sampled based on the second sampling direction and the constraint conditions, so that the atom type approximates the conditional distribution that satisfies the constraint conditions. The output module is used to output a complete molecular structure containing the known molecular fragment and capable of binding to the target protein pocket when the iterative updates of the continuous modal sampler and the discrete modal sampler both meet the preset termination condition.

[0065] Specifically, the pre-trained generative model module is responsible for learning the protein-ligand interaction rules at a global level and generating molecular structures that can specifically bind to the target protein pocket. The model takes the structural information of the target protein pocket as input and outputs the molecular 3D structure and atom types. The model progressively denoises and samples based on diffusion directions predicted by a neural network, achieving the generation process from random distribution to high-quality molecules. This model can automatically generate molecules with complete structures, reasonable conformations, and good binding affinity, providing a foundation for subsequent local optimization.

[0066] Secondly, the continuous modal sampler is used to handle continuous variables in the molecular structure, namely spatial coordinates. This module takes the continuous modal sampling direction output by the pre-trained model and the atomic coordinates of the existing molecular fragments as input, and comprehensively considers the global generation trend and conditional constraint information of the pre-trained model during the sampling process. Through dynamic correction of the sampling direction, the continuous modal sampler can effectively guide the molecular structure to tend towards a reasonable conformation in space, ensuring that the generated result is consistent with the given molecular fragment.

[0067] Finally, the discrete modal sampler is used to process discrete variable information in the molecular structure, mainly including features such as atom type and chemical bond type. This module takes the discrete modal sampling direction output by the pre-trained model and the atom type of the known molecular fragment as input, and corrects the sampling direction by combining conditional constraints, thereby achieving efficient sampling of discrete variables. The discrete modal sampler ensures the chemical rationality of the sampling results and the consistency with the given molecular fragment.

[0068] Since these systems and the principles underlying the problems they address are similar to the aforementioned plug-and-play lead compound optimization method, the implementation of this system can be found in the implementation of the aforementioned method, and the repetitions will not be repeated.

[0069] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0070] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in this invention may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A plug-and-play method for optimizing lead compounds, characterized in that, Includes the following steps: S1. Obtain target protein pocket information and known molecular fragment information, wherein the known molecular fragment information includes determined atomic coordinates and atomic types; fix the atomic coordinates and atomic types as constraints, and initialize the atomic coordinates and atomic types in the unknown molecular region to generate the initial molecular structure to be sampled; S2. Input the target protein pocket information and initial molecular structure into the pre-trained generation model to obtain the first sampling direction of the pre-trained generation model for continuous modal output and the second sampling direction for discrete modal output; S3. Based on the first sampling direction and the constraint conditions, the atomic coordinates of the molecule to be sampled are iteratively updated using a continuous modal sampler so that the atomic coordinates approximate the conditional distribution that satisfies the constraint conditions. S4. Based on the second sampling direction and the constraint conditions, the atom types of the molecule to be sampled are iteratively updated using a discrete modal sampler so that the atom types approximate the conditional distribution that satisfies the constraint conditions. S5. When the iterative updates of the continuous modal sampler and the discrete modal sampler both meet the preset termination condition, the complete molecular structure containing the known molecular fragment and capable of binding to the target protein pocket is output.

2. The method as described in claim 1, characterized in that, In step S1, fixing the atomic coordinates and atomic type as constraints specifically includes: The atomic coordinates and atom types of each atom contained in the known molecular fragment remain unchanged during the iterative update process, and the known molecular fragment serves as an immutable structural anchor point in the complete molecular structure.

3. The method as described in claim 1, characterized in that, In step S1, the atomic coordinates and atom types in the unknown molecular region are initialized, specifically including: The unknown molecular region is the region containing atoms other than those contained in the known molecular fragment in the complete molecular structure; The atomic coordinates of each atom in the unknown molecular region are randomly initialized using a Gaussian distribution; The atom types of each atom in the unknown molecular region are randomly initialized using a uniform distribution.

4. The method as described in claim 1, characterized in that, In step S2, the pre-trained generative model is a diffusion generative model trained under the framework of structure drug discovery. The model takes the three-dimensional structure of the protein pocket as the conditional input and outputs the three-dimensional molecular structure and atom type that can bind to the protein pocket. The first sampling direction is the noise prediction direction of the pre-trained generation model for the coordinates of each atom in the molecule to be sampled, and the second sampling direction is the probability prediction distribution of each atom type in the molecule to be sampled by the pre-trained generation model.

5. The method as described in claim 1, characterized in that, In step S3, the atomic coordinates of the molecule to be sampled are iteratively updated using a continuous modal sampler, specifically including: A continuous variational distribution is introduced for the atomic coordinates, which is used to approximate the true conditional distribution of the atomic coordinates under the constraints. Based on the first sampling direction output by the pre-trained generative model and the atomic coordinate constraints of the known molecular fragment, calculate the gradient of the divergence between the variational distribution and the true conditional distribution with respect to the variational distribution parameters. The parameters of the variational distribution are iteratively updated according to the gradient until a preset number of iterations is reached, so as to gradually correct the atomic coordinates.

6. The method as described in claim 5, characterized in that, The gradient formula for calculating the divergence between the variational distribution and the true conditional distribution with respect to the parameters of the variational distribution is as follows: in, This represents a pre-trained neural network. Represents the atomic coordinates in a complete molecule. Represents the atomic coordinates in a known molecular fragment. Represents a continuous diffusion model The latent variable corresponding to time. Represents distribution The parameters, The hyperparameters representing the control constraint weights The mapping matrix represents the generated molecules. Atoms and fragment structures contained The mapping relationship of the atoms contained.

7. The method as described in claim 1, characterized in that, In step S4, the atom types of the molecule to be sampled are iteratively updated using a discrete modal sampler, specifically including: A discrete variational distribution is introduced for the atom type, which is used to approximate the true conditional distribution of the atom type under the constraints. Based on the second sampling direction output by the pre-trained generative model and the atomic type constraints of the known molecular fragment, calculate the gradient of the divergence between the variational distribution and the true conditional distribution with respect to the variational distribution parameters; The parameters of the variational distribution are iteratively updated according to the gradient until a preset number of iterations is reached, so as to achieve gradual correction of the atom type.

8. The method as described in claim 7, characterized in that, The gradient formula for calculating the divergence between the variational distribution and the true conditional distribution with respect to the parameters of the variational distribution is as follows: in, Indicates the types of atoms in a complete molecule. Indicates the types of atoms in a known molecular fragment. Indicates the first The probability parameters of each category, Representing the discrete diffusion model The latent variable corresponding to time. This represents the latent variable at time 0 in the discrete diffusion model, i.e., the clean sample without noise. This means that all element values ​​are 0. The vector, This represents the forward process of the diffusion model. This represents the reverse process of the unconditional diffusion model. This represents the inverse process of the diffusion model corresponding to the variational distribution. This represents the marginal distribution of the unconditional diffusion model. Indicates the variational distribution Exported implicit variables The marginal distribution that is satisfied.

9. The method as described in claim 1, characterized in that, In step S5, the preset termination condition is that the number of iterations reaches a preset number of steps. The preset number of steps is the same as the total number of denoising steps used by the pre-trained generative model during the training phase, and the learning rate for each iteration is set to 0.

1.

10. A plug-and-play lead compound optimization system, characterized in that, The system comprises: An initialization module is used to acquire target protein pocket information and known molecular fragment information, wherein the known molecular fragment information includes determined atomic coordinates and atomic types; the atomic coordinates and atomic types are fixed as constraints, and the atomic coordinates and atomic types in the unknown molecular region are initialized to generate the initial molecular structure to be sampled; The pre-trained generative model module is used to input the target protein pocket information and initial molecular structure into the pre-trained generative model, and obtain the first sampling direction of the pre-trained generative model for continuous modal output and the second sampling direction for discrete modal output; A continuous modal sampler is used to iteratively update the atomic coordinates of the molecule to be sampled based on the first sampling direction and the constraint conditions, so that the atomic coordinates approximate the conditional distribution that satisfies the constraint conditions. A discrete modal sampler is used to iteratively update the atom type of the molecule to be sampled based on the second sampling direction and the constraint conditions, so that the atom type approximates the conditional distribution that satisfies the constraint conditions. The output module is used to output a complete molecular structure containing the known molecular fragment and capable of binding to the target protein pocket when the iterative updates of the continuous modal sampler and the discrete modal sampler both meet the preset termination condition.

Images