Method for directed optimization of ligand molecule generation towards target-aware molecule diffusion model

Abstract

This paper presents a method for targeted optimization of ligand molecule generation based on a target-aware molecular diffusion model, belonging to the field of computer-aided drug design technology. It addresses the shortcomings of existing ligand molecule generation methods, such as the inability to target molecular properties, the lack of evaluability and correction of intermediate generation trajectories, and the poor flexibility due to reliance on retraining. The method comprises the following stages: Preparation: A pre-trained model is used to generate ligand molecules and intermediate state samples, and each sample is labeled with molecular attributes (high-quality / low-quality); Training: A binary classifier for molecular attributes is trained based on the samples and labels; Inference: During molecule generation, the trained classifier is introduced to evaluate the attributes of intermediate states. The weighted evaluation is compared with a threshold and a set backstep number to determine whether to trigger a backstep operation, thereby guiding the model to continue generating from a better intermediate state, ultimately obtaining a ligand molecule with the desired attributes. This achieves targeted optimization of single or multiple molecular attributes. Its main application is in computer-aided drug design.

Description

Technical Field

[0001] This invention relates to the field of computer-aided drug design technology. Background Technology

[0002] Structure-based drug design (SBDD) is a core direction in the field of drug development. Its core objective is to generate highly specific and druggable ligand molecules (i.e., "keys") for protein pockets (three-dimensional spatial regions on protein pockets that can bind to drug ligands, i.e., "locks"). Deep generation methods, represented by diffusion models, have made significant progress in ligand molecule generation, but still have insurmountable shortcomings in practical drug development scenarios.

[0003] Existing diffusion-based ligand molecule generation methods are based on a given time... Below, based on protein information From initial molecular information Step-by-step generation to final molecular information Initial molecular information It is based on protein information The generated random data, each step based on protein information Molecular information ,time As input, predict the molecular information for the next step. .

[0004] Existing diffusion-based ligand molecule generation methods suffer from three core problems:

[0005] 1. Lack of guidance, unable to achieve targeted optimization: Existing diffusion model-based generation methods have strong randomness in each step of the prediction, lacking specific guidance for the generation direction at each step. The generation direction is uncontrollable and cannot be targeted to optimize molecular properties. Existing diffusion models rely solely on protein pocket conditions during molecule generation, and each step of denoising prediction is highly random, lacking precise guidance based on attributes. It is impossible to perform targeted optimization for key drug properties such as binding affinity, drug similarity (QED), and synthetic feasibility (SA), and it is impossible to perform targeted optimization for the specific properties of the final molecule. The generated molecules are difficult to meet the actual needs of drug development, and a large amount of intermediate molecule information does not belong to molecules that conform to chemical rules, so it is impossible to calculate the specific properties of intermediate molecule information.

[0006] 2. Inability to identify and correct intermediate trajectories: The intermediate generation trajectories cannot be evaluated or corrected. During the diffusion process, a large number of intermediate molecules do not meet the chemical rationality requirements, making it impossible to calculate their properties or judge the quality of the trajectories. The model can only proceed according to fixed steps and cannot identify inferior trajectories and correct them in time, resulting in poor final molecule properties and a high rate of invalid generation.

[0007] 3. Optimization of ligand molecules relies on retraining, which is costly and inflexible: Optimization relies on retraining, which is costly and inflexible. Most existing molecular property optimization methods require retraining the basic diffusion model, which is time-consuming and computationally expensive. It is impossible to perform fast optimization directly on the already trained target perception model, making it difficult to adapt to the flexible adjustment needs of multiple targets and multiple scenarios in drug development.

[0008] Therefore, the above problems urgently need to be solved. Summary of the Invention

[0009] The purpose of this invention is to address the problems of existing methods for generating ligand molecules, such as the inability to target and optimize molecular properties, the lack of evaluability and correction of intermediate generation trajectories, and the poor flexibility of optimization due to reliance on retraining. This invention provides a method for targeted optimization of ligand molecule generation based on a target-aware molecular diffusion model.

[0010] Methods for targeted optimization of ligand molecule generation based on target-sensing molecular diffusion models include:

[0011] Preparation phase: The pre-trained target-aware molecular diffusion model acts as a molecular generator, taking protein information as input and performing multiple molecular predictions in time-division. Each molecular prediction generates the corresponding final ligand molecular information. Using the final ligand molecular information and the intermediate ligand molecular information generated at each time step during the molecular prediction process, a set of training samples is constructed, and molecular attribute labels are defined for each training sample in all sets of training samples corresponding to the same protein information.

[0012] Training phase: The corresponding molecular attribute binary classifier is trained using each training sample and its corresponding molecular attribute label. The protein information and molecular information at each time step in each training sample are used as the input data of the corresponding molecular attribute binary classifier, and the molecular attribute label corresponding to the training sample is used as the ground truth of the corresponding molecular attribute binary classifier. The molecular attribute label includes high quality and poor quality. The molecular attributes of the binary classifier are the same as the molecular attributes of the label of its corresponding training sample.

[0013] Inference Phase: The pre-trained target-aware molecular diffusion model performs molecular prediction based on current protein information. During molecular prediction, a binary classifier for the corresponding molecular attribute type is introduced, representing one or more molecular attributes to be optimized. This process is initiated from the time step... In the direction towards time step 0, the current protein information and the intermediate ligand molecule information from each time step in the molecular prediction process are input as a sample to be optimized into the input of the various types of molecular attribute binary classifiers, and the attribute values ​​of the corresponding molecular attributes are evaluated; for the current time step The weighted sum of all corresponding evaluation results yields a joint evaluation result. The final evaluation result is determined based on whether the joint evaluation result exceeds a threshold and whether it exceeds the preset total rollback time steps. The system determines whether a rollback operation has been triggered. If a rollback operation is triggered and executed, the pre-trained target-sensing molecular diffusion model adjusts its time step accordingly. Information on intermediate ligand molecules The intermediate ligand molecule information is updated and molecules are predicted until the final ligand molecule information is output at time step 0, thereby achieving targeted optimization of ligand molecule information with single or multiple molecular properties.

[0014] Preferably, the molecular properties include binding affinity, drug similarity (QED), and synthetic accessibility (SA).

[0015] Preferably, the implementation method for defining the molecular attribute labels of each training sample in all training samples belonging to the same protein information is as follows:

[0016] In a set of training samples generated by evaluating the same protein information for a single molecular prediction The corresponding training samples The final ligand molecule information belongs to the attribute value of a certain molecular property, and is sorted from best to worst according to the superiority or inferiority of the attribute value. The first few are listed. The information of the final ligand molecule and its corresponding intermediate ligand molecule indicates that this type of molecule is tagged as high-quality and ranked lower. The information of the final ligand molecule and its corresponding intermediate ligand molecule indicates that the molecular attribute label is poor; among them, The total number of training samples generated for a single molecular prediction corresponding to information belonging to the same protein.

[0017] Preferably, the training method for the binary classifiers of various molecular attributes during the training phase is as follows:

[0018] The protein information and one molecule information in each training sample are used as the input to the molecular attribute binary classifier of the current category, and the molecular attribute label corresponding to the training sample is used as the output of the molecular attribute binary classifier of the current category. The molecular attribute binary classifier of the current category is then trained.

[0019] Each training sample includes protein information and molecular information, where the molecular information is either intermediate ligand molecule information or final ligand molecule information.

[0020] Preferably, the implementation method for determining whether to perform a rollback operation based on whether the joint evaluation result exceeds the threshold and the set constraints is as follows:

[0021] First, determine whether the joint evaluation result exceeds the threshold. If the result is yes, it indicates that the current time step... Information on intermediate ligand molecules To ensure high quality, no rollback operation is performed. If the result is negative, then it is checked whether the number of time steps between the current time step and the last time step that triggered and executed the rollback operation exceeds the preset rollback window. If the result is positive, it indicates that the current time step... Information on intermediate ligand molecules If the result is low quality, a rollback operation will be performed; if the result is negative, no rollback operation will be performed.

[0022] Preferably, after the rollback operation is triggered and executed, the pre-trained target-sensing molecular diffusion model adjusts according to the time step to which it has rolled back. Information on intermediate ligand molecules The method for updating intermediate ligand molecule information and predicting molecules until the final ligand molecule information is output at time step 0 is as follows:

[0023] At the current time step After triggering the rollback operation, the current time step Rewind preset total rewind time steps Then, rewind to the previous time step. This will retrieve the current protein information and allow you to revert to the previous time step. Information on intermediate ligand molecules As input to the pre-trained target-sensing molecular diffusion model, molecular predictions are continued, gradually obtaining the backtracked time step. Information on intermediate ligand molecules ;

[0024] Using Monte Carlo methods to combine information from the intermediate ligand molecules after regression Simulate the diffusion process to obtain the time step Next Single-objective reward based on molecular attributes ;

[0025] Single-objective reward weights that combine various predefined molecular attributes and single-target rewards The result after weighted summation is used as the time step. Total target reward ;

[0026] Information on the returned intermediate ligand molecules Total Target Rewards Superimpose the results to obtain the time step after the rollback. Information on intermediate ligand molecules ;

[0027] Combine current protein information and returned intermediate ligand molecule information. As a whole, each molecular attribute is input into a binary classifier for the various types of molecular attributes, and the attribute values ​​of the corresponding types of molecular attributes are evaluated. The weighted sum of all evaluation results is then obtained to obtain the time step after the rollback. The joint evaluation results are as follows;

[0028] Compare the time steps before and after rollback The joint evaluation results are ranked, and the intermediate ligand molecule information corresponding to the joint evaluation result with the higher value is used as the updated time step. Information on intermediate ligand molecules ;

[0029] Use the current protein information and the updated time step Information on intermediate ligand molecules The input is fed into the pre-trained target-aware molecular diffusion model for subsequent molecular prediction steps until the final ligand molecule information is output at time step 0, thereby achieving targeted optimization of ligand molecule information with single or multiple molecular properties.

[0030] The beneficial effects of this invention are:

[0031] This invention provides a method for targeted optimization of ligand molecule generation for target-aware molecular diffusion models. During the molecular prediction process of the target-aware molecular diffusion model in the inference stage, the intermediate ligand molecule information is evaluated and rewarded to guide correction, thereby achieving continuous numerical optimization of molecular properties and multi-objective collaborative optimization. This improves the binding affinity of ligand molecules while simultaneously optimizing drug similarity (QED) and synthetic feasibility (SA).

[0032] This invention achieves lightweight and efficient optimization by directly embedding an evaluation-optimization framework in the inference stage, without modifying or retraining the pre-trained target-aware molecular diffusion model. This significantly reduces computational and time costs and can quickly and accurately improve molecular properties on existing models, adapting to the rapid iteration needs of drug development.

[0033] Precisely oriented generation significantly enhances key drug properties. By using a binary classifier to evaluate intermediate molecules in real time at each step, the generation direction is guided by attributes for the noise reduction process. This can significantly improve the binding affinity between molecules and protein pockets, while optimizing drug similarity (QED) and synthetic feasibility (SA), thus achieving precise optimization of attribute values.

[0034] It supports intermediate trajectory evaluation and backtracking correction to avoid the accumulation of inferior quality. The pioneering evaluation-backtracking mechanism can identify intermediate generation trajectories that do not meet expectations in real time. Through backtracking and Monte Carlo optimization, it corrects deviations in a timely manner, blocks the transmission of inferior states, and significantly improves the chemical rationality and property excellence rate of generated molecules.

[0035] It supports multi-objective weighted collaborative optimization and is adapted to real drug design. Through multi-attribute reward weighted fusion, it can flexibly set the weights of objectives such as affinity, QED, and SA to achieve synergistic optimization of multiple drug-like attributes, avoid the overall decline in drug-likeness caused by single-objective optimization, and better fit the actual drug molecule development scenario. Attached Figure Description

[0036] Figure 1 This is a schematic diagram illustrating the principle of triggering and executing a rollback operation during the inference phase of the present invention.

[0037] Figure 2 This is the original diagram illustrating the principle for determining the quality of information about intermediate ligand molecules;

[0038] Figure 3 This is a comparison diagram showing whether the number of time steps between the current time step and the time step of the last time rollback operation was triggered exceeds the preset rollback window. Detailed Implementation

[0039] 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.

[0040] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0041] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0042] Specific Implementation Method 1: Combination Figure 1 and Figure 3 This embodiment describes a method for targeted optimization of ligand molecule generation based on a target-sensing molecular diffusion model, comprising:

[0043] Preparation phase: The pre-trained target-aware molecular diffusion model acts as a molecular generator, taking protein information as input and performing multiple molecular predictions in time-division. Each molecular prediction generates the corresponding final ligand molecular information. Using the final ligand molecular information and the intermediate ligand molecular information generated at each time step during the molecular prediction process, a set of training samples is constructed, and molecular attribute labels are defined for each training sample in all sets of training samples corresponding to the same protein information.

[0044] Training phase: The corresponding molecular attribute binary classifier is trained using each training sample and its corresponding molecular attribute label. The protein information and molecular information at each time step in each training sample are used as the input data of the corresponding molecular attribute binary classifier, and the molecular attribute label corresponding to the training sample is used as the ground truth of the corresponding molecular attribute binary classifier. The molecular attribute label includes high quality and poor quality. The molecular attributes of the binary classifier are the same as the molecular attributes of the label of its corresponding training sample.

[0045] Inference Phase: The pre-trained target-aware molecular diffusion model performs molecular prediction based on current protein information. During molecular prediction, a binary classifier for the corresponding molecular attribute type is introduced, representing one or more molecular attributes to be optimized. This process is initiated from the time step... In the direction towards time step 0, the current protein information and the intermediate ligand molecule information from each time step in the molecular prediction process are input as a sample to be optimized into the input of the various types of molecular attribute binary classifiers, and the attribute values ​​of the corresponding molecular attributes are evaluated; for the current time step The weighted sum of all corresponding evaluation results yields a joint evaluation result. The final evaluation result is determined based on whether the joint evaluation result exceeds a threshold and whether it exceeds the preset total rollback time steps. The system determines whether a rollback operation has been triggered. If a rollback operation is triggered and executed, the pre-trained target-sensing molecular diffusion model adjusts its time step accordingly. Information on intermediate ligand molecules The intermediate ligand molecule information is updated and molecules are predicted until the final ligand molecule information is output at time step 0, thereby achieving targeted optimization of ligand molecule information with single or multiple molecular properties. The value of is generally a constant of 1000. .

[0046] The inputs of the various molecular attribute binary classifiers are used to evaluate the attribute values ​​of the corresponding molecular attributes, so that the binary classifier outputs the evaluation result, which is the probability that the attribute value of the corresponding molecular attribute belongs to the high-quality category.

[0047] This implementation achieves targeted optimization generation of ligand molecules with desired molecular properties through the synergistic effect of a pre-trained target-aware molecular diffusion model and a molecular property binary classifier.

[0048] In the preparation phase, a training dataset is constructed using a large number of positive and negative samples generated by the pre-trained model and their intermediate information at multiple time steps. Each sample is assigned a binary classification attribute label of "high quality" or "poor quality", thus laying the foundation for subsequent attribute-guided training.

[0049] During the training phase, the corresponding molecular attribute binary classifier is trained based on these labeled samples, enabling the model to accurately distinguish the superior and inferior performance of ligand molecules on specific attributes at different time steps.

[0050] During the inference phase, given protein information, the pre-trained diffusion model incorporates one or more pre-trained binary classifiers during generation. These classifiers evaluate the properties of intermediate ligand molecules at each time step in real time and sum the results with weights to obtain a joint evaluation. The model then determines whether to trigger a backoff operation based on whether the result meets the backoff judgment mechanism (whether it exceeds a threshold and a preset backoff step count). If triggered, it backoffs to an earlier time step, updates the molecular information, and regenerates. This process iterates until a ligand molecule that meets one or more molecular property requirements is finally output. This significantly improves the directionality and controllability of ligand molecule generation, making it more suitable for actual drug design needs in target-aware scenarios. In specific applications, molecular property types include binding affinity, drug similarity (QED), and synthetic accessibility (SA).

[0051] Furthermore, the implementation method for defining the molecular attribute labels of each training sample in all training samples belonging to the same protein information is as follows:

[0052] In a set of training samples generated by evaluating the same protein information for a single molecular prediction The corresponding training samples The final ligand molecule information belongs to the attribute value of a certain molecular property, and is sorted from best to worst according to the superiority or inferiority of the attribute value. The first few are listed. The information of the final ligand molecule and its corresponding intermediate ligand molecule indicates that this type of molecule is tagged as high-quality and ranked lower. The information of the final ligand molecule and its corresponding intermediate ligand molecule indicates that the molecular attribute label is poor; among them, The total number of training samples generated for a single molecular prediction corresponding to information belonging to the same protein.

[0053] This preferred embodiment proposes an automatic attribute labeling method based on relative ranking. It evaluates the specific attribute values ​​of all final ligand molecules in a given molecular prediction for a specific molecular attribute and ranks them from best to worst. The top-ranked final ligand molecules and all their corresponding intermediate ligand molecules are labeled "high-quality," while the bottom-ranked final ligand molecules and their corresponding intermediate ligand molecules are labeled "low-quality." This method does not rely on pre-labeled attribute thresholds or manual judgment. Instead, it uses the relative quality comparison within the same protein information sample to define the label, effectively avoiding subjectivity and inconsistency in labeling. This ensures the discriminative power and balance between high-quality and low-quality categories in the training samples. Furthermore, it allows the binary classifier to learn subtle differences between molecules under the same attribute dimension, improving the discriminative power and stability of subsequent attribute-guided methods and providing a more reliable and self-consistent training supervision signal for the targeted optimization of the diffusion model.

[0054] Furthermore, during the training phase, the implementation method for training binary classifiers for various molecular attributes is as follows:

[0055] The protein information and one molecule information in each training sample are used as the input to the molecular attribute binary classifier of the current category, and the molecular attribute label corresponding to the training sample is used as the output of the molecular attribute binary classifier of the current category. The molecular attribute binary classifier of the current category is then trained.

[0056] Each training sample includes protein information and molecular information, where the molecular information is either intermediate ligand molecule information or final ligand molecule information.

[0057] In this preferred embodiment, when training various molecular attribute binary classifiers, the protein information and single molecule information (which can be the intermediate ligand molecule information in the diffusion process or the final ligand molecule information) in each training sample are used as input, and the "good" or "poor" attribute label corresponding to the sample is used as the output ground truth to conduct supervised training of the binary classifier.

[0058] Because the training samples extensively cover the molecular states throughout the complete generation trajectory from intermediate steps to the final step, the binary classifier can not only discriminate the attributes of the final generated ligand molecules but also effectively evaluate the attributes of intermediate ligand molecules at any time step during the generation process. This design greatly enhances the fine-grained control over the molecular generation process, providing a reliable basis for real-time evaluation of the attributes of intermediate molecules at each time step during the inference phase and triggering backtracking operations. Furthermore, the inclusion of protein information in the input ensures that the attribute evaluation has target-aware context specificity, thereby improving the overall accuracy of targeted optimization and the stability of the generation process.

[0059] Further, see Figure 2 The implementation method for determining whether to perform a rollback operation based on whether the joint evaluation results exceed the threshold and the set constraints is as follows:

[0060] First, determine whether the joint evaluation result exceeds the threshold. If the result is yes, it indicates that the current time step... Information on intermediate ligand molecules To ensure high quality, no rollback operation is performed. If the result is negative, then it is checked whether the number of time steps between the current time step and the last rolledback time step exceeds the preset rollback window. If the result is yes, it indicates that the current time step... Information on intermediate ligand molecules If the result is low quality, a rollback operation will be performed; if the result is negative, no rollback operation will be performed.

[0061] This preferred embodiment provides a rollback judgment mechanism based on dual constraints of joint evaluation results and rollback window: First, it is determined whether the joint evaluation result at the current time step exceeds a threshold. If it does, the intermediate ligand molecule information is considered superior (i.e., high quality), and generation continues directly without rollback. If the threshold is not exceeded, it is further checked whether the step difference between the current time step and the time step to which the rollback was last occurred exceeds a preset rollback window. Rollback is only performed if the window is not exceeded; otherwise, rollback is not performed even if the joint evaluation result is poor. This mechanism effectively avoids the problem of infinite loops or low generation efficiency caused by repeated rollbacks in continuously poor regions. The rollback window constraint limits the frequency and range of rollbacks, ensuring the convergence and computational efficiency of the generation process. At the same time, the threshold judgment ensures that only truly inferior intermediate states trigger rollback, avoiding unnecessary backtracking operations. Thus, a good balance between generation quality and operating efficiency is achieved while optimizing molecular properties in a targeted manner.

[0062] Furthermore, after the rollback operation is triggered and executed, the pre-trained target-sensing molecular diffusion model adjusts according to the time step to which it has rolled back. Information on intermediate ligand molecules The method for updating intermediate ligand molecule information and predicting molecules until the final ligand molecule information is output at time step 0 is as follows:

[0063] At the current time step After triggering the rollback operation, the current time step Rewind preset total rewind time steps Then, rewind to the previous time step. This will retrieve the current protein information and allow you to revert to the previous time step. Information on intermediate ligand molecules As input to the pre-trained target-sensing molecular diffusion model, molecular predictions are continued, gradually obtaining the backtracked time step. Information on intermediate ligand molecules ;

[0064] Using Monte Carlo methods to combine information from the intermediate ligand molecules after regression Simulate the diffusion process to obtain the time step Next Single-objective reward based on molecular attributes ;

[0065] Single-objective reward weights that combine various predefined molecular attributes and single-target rewards The result after weighted summation is used as the time step. Total target reward ;

[0066] Information on the returned intermediate ligand molecules Total Target Rewards Superimpose the results to obtain the time step after the rollback. Information on intermediate ligand molecules ;

[0067] Combine current protein information and returned intermediate ligand molecule information. The values ​​are input into the various classifiers for molecular attributes, and attribute values ​​for each class are evaluated. All evaluation results are then weighted and summed to obtain the time step after the rollback. The joint evaluation results are as follows;

[0068] Compare the time steps before and after rollback The joint evaluation results are ranked, and the intermediate ligand molecule information corresponding to the joint evaluation result with the higher value is used as the updated time step. Information on intermediate ligand molecules ;

[0069] Use the current protein information and the updated time step Information on intermediate ligand molecules As a whole, the ligand is input into the pre-trained target-aware molecular diffusion model for subsequent molecular prediction steps until the final ligand molecule information is output at time step 0, thereby achieving targeted optimization of ligand molecule information with single or multiple molecular properties.

[0070] In this preferred embodiment, after triggering the rollback operation, the joint evaluation results at each time step do not need to be evaluated during the rollback process. Instead, a refined molecular information update and optimization process is executed: First, the current time step is rolled back to an earlier time step according to the preset total number of rollback time steps, and a new state is predicted based on the intermediate ligand molecule information after the rollback. Then, the Monte Carlo method is introduced to simulate the diffusion process to calculate the single-target reward, and a weighted sum is performed by combining the preset weights of various molecular attributes to obtain the total target reward. This reward is superimposed with the intermediate ligand molecule information after the rollback, thereby generating new intermediate molecule information after attribute-guided enhancement. Next, the new information is jointly evaluated by the various molecular attribute binary classifiers used, and compared with the joint evaluation results at the same time step before the rollback. The intermediate ligand molecule information corresponding to the higher evaluation result is selected as the final updated state of that time step. Finally, subsequent molecular predictions are continued based on the updated state until the final ligand molecule is output. This approach significantly improves the quality of molecular information after backtracking by organically combining multiple operations such as backtracking, Monte Carlo reward estimation, weighted multi-objective fusion, merit-disadvantage comparison and optimal update. It enables the model to actively correct poor trends and select better intermediate states during the generation process, thereby achieving more stable and accurate targeted optimization of single or multiple molecular properties while ensuring molecular validity.

[0071] Verification experiment:

[0072] Taking the affinity-corresponding index VinaScore as an example, the results obtained by using a binary classifier trained with VinaScore to participate in the diffusion process of two diffusion models (TargetDiff, IPDiff) are shown in Table 1:

[0073] Table 1 shows the results of ligand molecule generation based on VinaScore-directed optimization.

[0074]

[0075] In Table 1, Mean represents the arithmetic mean of the indicator across all generated molecules, reflecting the overall average level. Median represents the median value of the indicator across all generated molecules, avoiding interference from extreme values ​​and reflecting robustness. VinaScore is the binding affinity score between the molecule and the protein pocket; a smaller (more negative) value indicates stronger binding affinity and higher drug potential. VinaMin is the optimal binding affinity value among all generated molecules; a smaller (more negative) value indicates stronger binding affinity of the optimal molecule.

[0076] As shown in Table 1, when the method of this invention is applied to the TargetDiff and IPDiff diffusion models respectively, the core optimization attributes combined with affinity (VinaScore, VinaMin) are significantly improved. At the same time, there are only slight fluctuations in non-core optimization attributes such as drug similarity (QED) and synthetic feasibility (SA), with minimal overall loss and no significant impact on molecular drugability. This fully verifies that the method can accurately target and enhance key optimization objectives while taking into account other drug properties. As shown in Table 1, after applying the method of this invention to the TargetDiff and IPDiff diffusion models respectively, the core optimization attributes combined with affinity (VinaScore and VinaMin) were significantly improved. Specifically, in the TargetDiff model, the mean VinaScore was optimized from -5.36 to -6.52, the median from -5.95 to -6.43, and the mean VinaMin from -6.46 to -6.96, with the median from -6.34 to -6.99. In the IPDiff model, the mean VinaScore was optimized from -6.73 to -7.79, the median from -6.98 to -7.75, and the mean VinaMin from -7.18 to -7.93, with the median from -7.04 to -7.88. Meanwhile, non-core optimization attributes such as drug similarity (QED) and synthetic feasibility (SA) showed only minor fluctuations, with minimal overall loss and no significant impact on druggability. Specifically, in the TargetDiff model, the mean and median of QED increased from 0.46 to 0.49, while the mean SA decreased slightly from 0.61 to 0.56, and the median remained unchanged at 0.60. In the IPDiff model, the mean and median of QED decreased slightly from 0.50 / 0.51 to 0.47, while the mean SA decreased from 0.61 to 0.59, and the median decreased from 0.61 to 0.58. This fully validates that the proposed method can precisely target and enhance key optimization objectives while considering other druggable attributes. It demonstrates stable and reliable optimization effects on two typical diffusion models, and the synergistic performance of multiple attributes better aligns with the actual needs of drug development.

[0077] Overall, the method of this invention can achieve precise targeted optimization of binding affinity with minimal sacrifice of non-target attributes, and the synergistic performance of multiple attributes is more in line with the actual needs of drug development. Although the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for targeted optimization of ligand molecule generation based on a target-sensing molecular diffusion model, characterized in that, include: Preparation phase: The pre-trained target-aware molecular diffusion model acts as a molecular generator, taking protein information as input and performing multiple molecular predictions in time-division. Each molecular prediction generates the corresponding final ligand molecular information. Using the final ligand molecular information and the intermediate ligand molecular information generated at each time step during the molecular prediction process, a set of training samples is constructed, and molecular attribute labels are defined for each training sample in all sets of training samples corresponding to the same protein information. Training phase: The corresponding molecular attribute binary classifier is trained using each training sample and its corresponding molecular attribute label. The protein information and molecular information at each time step in each training sample are used as the input data of the corresponding molecular attribute binary classifier, and the molecular attribute label corresponding to the training sample is used as the ground truth of the corresponding molecular attribute binary classifier. The molecular attribute label includes high quality and poor quality. The molecular attributes of the binary classifier are the same as the molecular attributes of the label of its corresponding training sample. Inference Phase: The pre-trained target-aware molecular diffusion model performs molecular prediction based on current protein information. During molecular prediction, a binary classifier for the corresponding molecular attribute type is introduced, representing one or more molecular attributes to be optimized. This process is initiated from the time step... In the direction towards time step 0, the current protein information and the intermediate ligand molecule information from each time step in the molecular prediction process are input as a sample to be optimized into the input of the various types of molecular attribute binary classifiers, and the attribute values ​​of the corresponding molecular attributes are evaluated; for the current time step The weighted sum of all corresponding evaluation results yields a joint evaluation result. The final evaluation result is determined based on whether the joint evaluation result exceeds a threshold and whether it exceeds the preset total rollback time steps. The system determines whether a rollback operation has been triggered. If a rollback operation is triggered and executed, the pre-trained target-sensing molecular diffusion model adjusts its time step accordingly. Information on intermediate ligand molecules The intermediate ligand molecule information is updated and molecules are predicted until the final ligand molecule information is output at time step 0, thereby achieving targeted optimization of ligand molecule information with single or multiple molecular properties.

2. The method for targeted optimization of ligand molecule generation based on a target-sensing molecular diffusion model according to claim 1, characterized in that, Molecular properties include binding affinity, drug similarity (QED), and synthetic accessibility (SA).

3. The method for targeted optimization of ligand molecule generation based on a target-sensing molecular diffusion model according to claim 1, characterized in that, The implementation method for defining the molecular attribute labels of each training sample in all training samples belonging to the same protein information is as follows: In a set of training samples generated by evaluating the same protein information for a single molecular prediction The corresponding training samples The final ligand molecule information belongs to the attribute value of a certain molecular property, and is sorted from best to worst according to the superiority or inferiority of the attribute value. The first few are listed. The information of the final ligand molecule and its corresponding intermediate ligand molecule indicates that this type of molecule is tagged as high-quality and ranked lower. The information of the final ligand molecule and its corresponding intermediate ligand molecule indicates that the molecular attribute label is poor; among them, The total number of training samples generated for a single molecular prediction corresponding to information belonging to the same protein.

4. The method for targeted optimization of ligand molecule generation based on a target-sensing molecular diffusion model according to claim 1, characterized in that, The training phase involves training binary classifiers for various molecular attributes as follows: The protein information and one molecule information in each training sample are used as the input to the molecular attribute binary classifier of the current category, and the molecular attribute label corresponding to the training sample is used as the output of the molecular attribute binary classifier of the current category. The molecular attribute binary classifier of the current category is then trained. Each training sample includes protein information and molecular information, where the molecular information is either intermediate ligand molecule information or final ligand molecule information.

5. The method for targeted optimization of ligand molecule generation based on a target-sensing molecular diffusion model according to claim 1, characterized in that, The method for determining whether to perform a rollback operation based on whether the joint evaluation results exceed the threshold and the set constraints is as follows: First, determine whether the joint evaluation result exceeds the threshold. If the result is yes, it indicates that the current time step... Information on intermediate ligand molecules To ensure high quality, no rollback operation is performed. If the result is negative, then it is checked whether the number of time steps between the current time step and the last time step that triggered and executed the rollback operation exceeds the preset rollback window. If the result is positive, it indicates that the current time step... Information on intermediate ligand molecules If the result is low quality, a rollback operation will be performed; if the result is negative, no rollback operation will be performed.

6. The method for targeted optimization of ligand molecule generation based on a target-sensing molecular diffusion model according to claim 1, characterized in that, Once the rollback operation is triggered and executed, the pre-trained target-sensing molecular diffusion model will adjust according to the time step it has rolled back to. Information on intermediate ligand molecules The method for updating intermediate ligand molecule information and predicting molecules until the final ligand molecule information is output at time step 0 is as follows: At the current time step After triggering the rollback operation, the current time step Rewind preset total rewind time steps Then, rewind to the previous time step. This will retrieve the current protein information and allow you to revert to the previous time step. Information on intermediate ligand molecules As input to the pre-trained target-sensing molecular diffusion model, molecular predictions are continued, gradually obtaining the backtracked time step. Information on intermediate ligand molecules ; Using Monte Carlo methods to combine information from the intermediate ligand molecules after regression Simulate the diffusion process to obtain the time step Next Single-objective reward based on molecular attributes ; Single-objective reward weights that combine various predefined molecular attributes and single-target rewards The result after weighted summation is used as the time step. Total target reward ; Information on the returned intermediate ligand molecules Total Target Rewards Superimpose the results to obtain the time step after the rollback. Information on intermediate ligand molecules ; Current protein information and returned intermediate ligand molecule information As a whole, each molecular attribute is input into a binary classifier for the various types of molecular attributes, and the attribute values ​​of the corresponding types of molecular attributes are evaluated. The weighted sum of all evaluation results is then obtained to obtain the time step after the rollback. The joint evaluation results are as follows; Compare the time steps before and after rollback The joint evaluation results are ranked, and the intermediate ligand molecule information corresponding to the joint evaluation result with the higher value is used as the updated time step. Information on intermediate ligand molecules ; Use the current protein information and the updated time step Information on intermediate ligand molecules The input is fed into the pre-trained target-aware molecular diffusion model for subsequent molecular prediction steps until the final ligand molecule information is output at time step 0, thereby achieving targeted optimization of ligand molecule information with single or multiple molecular properties.

Images