Artificial intelligence-based molecular characterization and soft-skeleton-constrained drug discovery system

Abstract

This application designs an artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system, comprising: a data collection module, a sub-scaffold analysis module, a 3D simulated molecular docking module, a Bayesian ridge regressor training module, a VAE generative model pre-training module, a reinforcement learning module, and a sampling evaluation module. By using sub-scaffold analysis and 3D simulated molecular docking information, this application can better understand and utilize the complexity and diversity of chemical space. Furthermore, by using a Bayesian ridge regressor and a VAE generative model, this application can generate novel, rationally structured, and potentially active drug molecules. Therefore, this artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system, by combining deep learning technology, cheminformatics, drug design, and computer-aided drug design methods, can significantly improve the efficiency and success rate of drug discovery.

Description

Technical Field

[0001] This invention belongs to the field of drug development, specifically involving an artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system. Background Technology

[0002] Drug discovery is a costly, high-risk, and long-term process, and deep learning based on artificial intelligence has opened up new avenues for this field. Generative models can help researchers explore novel and diverse chemical structures from a vast chemical space, but this approach is still immature. This is mainly because most models can only handle single molecular characterization methods such as SMILES strings or graph structures, and the molecular generation process is highly random and uninterpretable. Using rigid sub-backbone constraints makes it difficult for models to learn the formation process of the sub-backbone and understand the drug mechanism. Molecules can be characterized in various ways: SMILES strings, molecular fingerprints, graph structures, or three-dimensional structures. Currently, none of these characterization methods are perfect and cannot capture the chemical and physical information of different types of molecules.

[0003] Most existing machine learning-assisted drug discovery methods are based on molecular design and screening methods such as property-based, target-based, and ligand structure-based approaches. These are entirely data-driven, and the training process and parameters within the models are irregular and uninterpretable, making it difficult to guide the training process with expert knowledge and experience. This approach relies on ligand information during training and is therefore susceptible to biases in the training dataset related to ligands. Furthermore, the molecular characterization methods input during training are limited to single SMILES molecular formulas or graphical structures, failing to utilize the 3D docking information between the binding pocket and the ligand. Additionally, some structure-based molecular generation methods impose rigid constraints, such as determining the existence of the backbone or rigidly fixing the target backbone during generation, without a flexible, multi-gradient backbone generation path, understanding the composition and importance of different functional groups, or a judgment process. Therefore, the molecular generation model cannot understand the formation process and principles of the backbone, resulting in molecular structures lacking diversity and rationality.

[0004] The main drawback of existing technologies is their inability to effectively integrate multiple characterization information of molecules, such as smiles, molecular fingerprints, and 3D structures. This results in incomplete capture of the chemical and physical properties of molecules by the models, leading to low efficiency and quality in generating molecular structures. Another drawback is that existing technologies do not use soft sub-skeleton constraints to train the generative models. If hard structure-based molecular generation methods are used, an effective gradient descent process cannot be provided for the models, making it difficult for them to understand the constituent elements of the skeleton, resulting in a lack of diversity and rationality in the generated molecular structures. Summary of the Invention

[0005] To address the aforementioned issues, this application designs an artificial intelligence-based molecular characterization and soft sub-scaffold constraint-based drug discovery system. The aim is to improve the efficiency and success rate of drug discovery by combining deep generative models, multi-molecule characterization, soft sub-scaffold constraints, and multi-task optimization of reward functions.

[0006] The AI-based molecular characterization and soft sub-scaffold-constrained drug discovery system includes: a data collection module, a sub-scaffold analysis module, a 3D simulated molecular docking module, a Bayesian ridge regressor training module, a VAE generative model pre-training module, a reinforcement learning module, and a sampling evaluation module.

[0007] The data collection module is used to collect a wide range of drug molecule datasets, target-related ligand molecule datasets, and crystal structure data of target protein binding to ligands.

[0008] The VAE generative model pre-training module is used to learn the basic syntax of SMILES molecular formulas and generate rich and effective chemical SMILES strings.

[0009] The sub-skeleton analysis module uses ScaffoldGraph and a Bayesian theory-based sub-skeleton analysis method to perform sub-skeleton analysis on target-related ligand molecule datasets to obtain information on one or more sub-skeletons that contribute most to the activity.

[0010] The 3D simulated molecular docking module performs 3D simulated docking on molecular samples in the target-related ligand molecule dataset and the crystal structure data of the target protein and ligand binding, to obtain the binding conformation and docking score containing molecular 3D information, as well as a dataset with binding energy.

[0011] The Bayesian Ridge Regressor Training Module trains the Bayesian Ridge Regressor to fit the 3D docking results of molecules by using the molecular fingerprinting method on molecular samples with docking scores, thereby obtaining a regressor that can quickly and accurately predict the docking scores of molecules.

[0012] The reinforcement learning module optimizes the VAE generative model pre-training module by combining the Bayesian Ridge Regressor training module with a multi-objective optimization reward function formed by combining soft sub-skeleton constraints.

[0013] The sampling evaluation module is used to test the VAE generation model generated by the reinforcement learning module. If it does not meet the requirements for drug design, the soft sub-scaffold constraints will be adjusted or the Bayesian ridge regressor training module will be adjusted.

[0014] Preferably, the VAE generative model pre-training module is used to learn the basic syntax of SMILES molecular formulas and generate rich and valid chemical SMILES strings; including:

[0015] The AE model was trained using a combination of reconstruction loss and KL divergence loss.

[0016] Each SMILES string is converted into a vector using embedding, and these vectors serve as input to the variational autoencoder model. The encoder layer of the variational autoencoder model learns how to transform the input vector into a 50-dimensional continuous latent space vector, and the decoder layer of the variational autoencoder model learns how to decode the original SMILES string from this 50-dimensional continuous latent space vector.

[0017] Preferably, when training the variational autoencoder model, the input includes not only the SMILES molecular formula, but also basic molecular information to help the variational autoencoder model better understand the chemical space, thereby improving the quality and diversity of generated molecules.

[0018] Preferably, when performing sub-skeleton analysis on the target-related ligand molecule dataset, other sub-skeleton structures can be added manually.

[0019] Preferably, the 3D simulated molecular docking module performs 3D simulated docking on molecular samples from the target-related ligand molecule dataset and the crystal structure data of the target protein and ligand binding, including:

[0020] The dataset with binding energy obtained from the 3D simulated molecular docking module is input. First, the ECFP molecular fingerprint of each sample in 2048 dimensions is calculated using rdkit. Then, the molecular fingerprint is used as the feature matrix and the binding energy is used as the target vector to fit the model using a Bayesian ridge regressor. The model loss function is the mean squared error plus the L2 regularization term determined by the Bayesian method.

[0021] Preferably, the reinforcement learning module optimizes the VAE generative model pre-training module by combining the Bayesian ridge regressor training module with a multi-objective optimization reward function formed by soft sub-skeleton constraints; including:

[0022] In each round of reinforcement learning, 500 latent space vectors of length 50 are sampled from the model, and these vectors are converted into 500 SMILES molecular expressions by the decoder.

[0023] Next, the efficiency and reward values ​​of these molecules are calculated, reflecting the quality and potential activity of the molecules;

[0024] Then, these reward values ​​are used to optimize the VAE generative model's sampling strategy in the latent space and the decoder layer weights.

[0025] Preferably, the multi-objective optimization reward function is composed of multiple objectives and different weight values, including soft sub-skeleton constraints and a Bayesian ridge regressor.

[0026] Preferably, gradient descent is used when optimizing the pre-training module of the VAE generative model, and the learning rate decay strategy is set to decrease by 10% every 50 rounds.

[0027] The advantages and effects of this application are as follows:

[0028] This application designs an artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system, comprising: a data collection module, a sub-scaffold analysis module, a 3D simulated molecular docking module, a Bayesian ridge regressor training module, a VAE generative model pre-training module, a reinforcement learning module, and a sampling evaluation module. By using sub-scaffold analysis and 3D simulated docking information, this application can better understand and utilize the complexity and diversity of chemical space. Furthermore, by using a Bayesian ridge regressor and a VAE generative model, this application can generate novel, rationally structured, and potentially active drug molecules. Therefore, this artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system, by combining deep learning technology, cheminformatics, drug design, and computer-aided drug design methods, can significantly improve the efficiency and success rate of drug discovery.

[0029] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0030] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0032] Figure 1 A flowchart of the AI-based molecular characterization and soft-framework-constrained drug discovery system designed for this application. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0034] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0035] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0036] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0037] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0038] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0039] Example

[0040] This embodiment mainly introduces the detailed design of an artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system. Please refer to... Figure 1 Artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery systems mainly include:

[0041] The module includes a data collection module, a sub-skeleton analysis module, a 3D simulated molecular docking module, a Bayesian ridge regressor training module, a VAE generative model pre-training module, a reinforcement learning module, and a sampling evaluation module.

[0042] The data collection module is used to collect a wide range of drug molecule datasets, target-related ligand molecule datasets, and crystal structure data of target protein binding to ligands.

[0043] The sub-skeleton analysis module uses ScaffoldGraph and a Bayesian theory-based sub-skeleton analysis method to perform sub-skeleton analysis on target-related ligand molecule datasets, and combines expert knowledge to obtain information on one or more sub-skeletons that contribute most to activity.

[0044] ScaffoldGraph can generate a set of possible sub-backbone structures for each target-related ligand molecule sample according to specific splitting rules. Combining the activity tags of the ligand molecule dataset, the sub-backbone analysis module analyzes the number of molecules and activity ratios contained in each sub-backbone structure, and uses Bayesian theory to infer the contribution of each sub-backbone structure to molecule activity. The sub-backbone analysis module can provide a reference for users and medicinal chemists based on the relationship between sub-backbone structures and molecule activity.

[0045] Building upon this foundation, expert chemists can select from these structures or add their own to introduce soft structure constraints during the training phase of the variational autoencoder (VAC) model. These constraints are multi-gradient and soft, rather than hard, because the method provides multiple lists of sub-skeletons with inclusion or parallel relationships and calculates rewards based on the percentage of skeleton completion rather than the total structure. Different pharmacophores within the sub-skeletons have varying importance, reflecting the drug mechanism underlying the sub-skeletons. This improves the VCC model's understanding of sub-skeleton structures and drug mechanisms, increases the flexibility of molecular structure generation, and makes gradient descent during VCC model training smoother. Users can combine expert knowledge during the reinforcement learning phase to select optimal soft sub-skeleton constraints, thereby increasing the probability of generating active molecular structures.

[0046] The 3D simulation molecular docking module will perform 3D simulation docking on molecular samples in the target-related ligand molecule dataset and the crystal structure data of target protein and ligand binding, to obtain the binding conformation and docking score containing molecular 3D information, as well as the dataset with binding energy.

[0047] In this process, the 3D simulation molecular docking module uses the molecule's 3D information to simulate and generate possible docking conformations between the molecule and the target protein pocket, and calculates their binding energy, thereby obtaining the affinity between the molecule and the target. Various tools can be used for the docking process, such as SMINA and Glide.

[0048] The Bayesian Ridge Regressor Training Module uses the molecular fingerprinting method to train the Bayesian Ridge Regressor to fit the 3D docking results of molecules with docking scores, resulting in a regressor that can quickly and accurately predict the docking scores of molecules.

[0049] In this step, a dataset with binding energies is input. First, rdkit is used to calculate a 2048-dimensional ECFP molecular fingerprint for each sample. Then, the molecular fingerprint is used as the feature matrix, and the binding energy as the target vector, which is fitted using a Bayesian ridge regressor. The model loss function is the mean squared error plus an L2 regularization term determined by the Bayesian method. Its advantages include automatic determination of regularization parameters, generation of probability outputs, good performance on small sample datasets, and excellent accuracy and computational speed in predicting molecular docking affinity.

[0050] The VAE generative model pre-training module uses a wide range of drug molecule datasets and basic molecular properties as prior knowledge to pre-train the VAE generative model, enabling the model to learn the basic syntax of SMILES molecular formulas and acquire the ability to generate rich and effective chemical molecule representations.

[0051] A variational autoencoder (VAE) model is a generative model that learns the latent distribution of data and then samples from this distribution to generate new data. In this step, the VAE model learns the latent distribution of drug molecule data, samples from this distribution, and then generates new drug molecule SMILES sequences after passing through a decoder.

[0052] The molecular dataset contains a large number of drug-like molecules, each represented as a SMILES string. SMILES is a widely used method for representing chemical molecules, using specific grammatical rules to describe the molecular structure. The goal of pre-training is to enable the variational autoencoder model to learn the basic syntax and intrinsic rules of molecular structure from these SMILES strings, and to generate efficient, diverse SMILES molecular formulas by sampling in a smaller-dimensional space. To achieve this goal, the first step is embedding, using a one-to-one dictionary to convert each SMILES string into a vector, which serves as the input to the variational autoencoder model. The task of the variational autoencoder model is to reconstruct the input vectors as closely as possible, i.e., to generate vectors that are very close to the input vectors.

[0053] During training, the model's encoder layer learns how to transform the input vector into a 50-dimensional continuous latent space vector, and the model's decoder layer learns how to decode the original SMILES string from this 50-dimensional continuous vector. This process helps the variational autoencoder model understand the basic syntax of the SMILES string and the intrinsic rules of its molecular structure.

[0054] In addition to the molecular formulas of SMILES, the input also includes basic molecular information such as molecular weight, qed, and logP. This information, as prior knowledge, helps the model better understand chemical space, thereby improving the quality and diversity of generated molecules.

[0055] When training the variational autoencoder model, a combination of reconstruction loss and KL divergence loss is used. The reconstruction loss ensures that the model can accurately reconstruct the input drug molecule, while the KL divergence loss makes the latent distribution learned by the model approximate the prior distribution.

[0056] After pre-training, the variational autoencoder model will possess the ability to generate rich and efficient representations of chemical molecules. This means that, given a 50-dimensional continuous vector, the variational autoencoder model can generate a corresponding string of SMILES, i.e., a possible molecular structure. This capability is crucial for subsequent drug molecule design, because the sampling rules of this 50-dimensional continuous vector can be optimized through reinforcement learning to search for molecular structures with desired properties.

[0057] The reinforcement learning module uses a multi-objective optimization reward function composed of soft sub-skeleton constraints and a Bayesian ridge regressor training module to perform reinforcement learning, resulting in a model that can sample and generate novel and reasonable drug molecule structures with potential active properties.

[0058] In the reinforcement learning step, a multi-objective optimization reward function is used, which combines soft sub-skeleton constraints and Bayesian ridge regressors. The goal of this reward function is to optimize the model so that it can generate novel, rational, and potentially active drug molecule structures.

[0059] In each round of reinforcement learning, 500 latent space vectors of length 50 are sampled from the model. These vectors are then converted into 500 SMILES molecular expressions by the decoder. Next, the efficiency and reward values ​​of these molecules are calculated, reflecting the quality and potential activity of the molecules.

[0060] Then, these reward values ​​are used to optimize the model's sampling strategy in the latent space and the decoder layer weights. The reward function consists of multiple objectives and different weight values, including soft sub-skeleton constraints, Bayesian ridge regressors, and other molecular property conditions. In this way, after multiple iterations, the model can generate molecules that perform better overall across various objective tasks.

[0061] Gradient descent was used during the optimization process. To prevent the optimization process from becoming unstable due to an excessively large learning rate, a learning rate decay strategy was implemented: the learning rate decreased by 10% every 50 epochs. This allows the model to quickly approach the optimization target in the early stages, while allowing for more precise adjustments in later stages, thereby improving the optimization effect.

[0062] Through this step, the VAE generation model can not only generate novel, rational, and potentially active drug molecules, but also understand the drug mechanism of ligand molecules binding to target proteins based on 3D simulation information of molecule-target docking and soft sub-scaffolds based on expert knowledge. This allows for optimization and adjustment of the VAE generation model, resulting in higher-quality drug molecules in subsequent generation processes.

[0063] The sampling evaluation module samples a batch of SMILES chemical formulas from the trained generative model. If these chemical formulas do not meet the requirements for drug design, the process returns to step 6 to adjust the reward function strategy, change the sub-skeleton constraints, or adjust the weights of various terms. Reinforcement learning is then applied to the generative model to optimize the sampling distribution in the latent space and the generated molecular results. If the generated molecules possess the desired properties and potential activity, the process terminates.

[0064] Furthermore, in addition to using ScaffoldGraph and Bayesian theory-based sub-skeleton analysis methods, other sub-skeleton analysis methods can be used, such as graph theory-based methods and machine learning-based methods.

[0065] Furthermore, in addition to using Smina for 3D simulation docking, other 3D simulation docking tools such as AutoDock and Glide can also be considered.

[0066] Furthermore, in addition to using the Bayesian ridge regression model to fit the molecular docking score, other machine learning regressor algorithms can be used as alternatives, such as linear regression, SVM, XGBoost, and neural networks.

[0067] Furthermore, in addition to using VAE as the framework for generative models, other generative models, such as Generative Adversarial Networks (GANs) and Transformers, can also be used.

[0068] Furthermore, in addition to molecular characterization methods such as SMILES, ECFP, and 3D, there are also molecular characterization methods such as selfies, graph structures, and different types of molecular fingerprints that can be used to help models for inference and prediction.

[0069] Drug discovery involves searching for drug molecular structures that can closely dock with and possess pharmacological activity against a specific protein target and pocket. Traditional drug development relies on extensive experimental verification, but the chemical space of drug molecules is vast, making it extremely time-consuming and labor-intensive to exhaustively search for all possible active structures experimentally—like finding a needle in a haystack. Artificial intelligence and deep learning play a more efficient role in this process, generating and screening molecular structures that are more likely to be active. However, another problem arises: chemical molecules are extremely complex systems, and enabling models to accurately recognize and understand molecular structures and their biochemical properties is crucial. Most generative models use SMILES strings or graph structures to represent molecular structures because string formats are closest to traditional natural language, and graph structures are closest to the ball-and-stick model of molecules. However, molecular descriptors contain very limited molecular information, such as the inability to express the 3D spatial structure and shape of molecules. Therefore, this invention combines multiple molecular descriptors, especially 3D molecular docking information, to help models better understand the process of molecule-protein docking.

[0070] Sub-scaffold constraints provide a pre-defined molecular scaffold structure for the generative model during the generation process. Based on these constraints, the model can generate active molecular structures that more closely resemble expert predictions. Most hard scaffold constraints are based on adding new branches and other structures to a fixed sub-scaffold. However, this method produces fixed molecular structures, and the model cannot understand the scaffold formation process, the importance of different pharmacophores, and their relationship with the drug's mechanism of action. Therefore, this invention proposes a soft scaffold constraint. By analyzing a ligand molecule library using a built-in ScaffoldGraph and Bayesian principles, the contribution of different scaffolds and functional groups to activity is obtained. A reward function is designed based on multiple lists of sub-scaffolds with inclusion or parallel relationships and the percentage of scaffold completion for the reinforcement learning phase. This helps the model better understand the relationship between the drug's mechanism of action and the scaffold constraints during training, generating more diverse and effective novel drug molecular structures.

[0071] This invention improves the efficiency and success rate of drug discovery by combining deep learning technology, cheminformatics, drug design, and computer-aided drug design (CADD). Specifically, by using sub-scaffold analysis and 3D simulation docking information, this application can better understand and utilize the complexity and diversity of chemical space. Furthermore, by using Bayesian ridge regressors and VAE generative models, this application can generate novel, rationally designed, and potentially active drug molecules.

[0072] The above description is merely a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter alterations to these embodiments within the spirit and principles of the present invention, achieved through conventional substitutions or by achieving the same function without departing from the principles and spirit of the present invention, fall within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system, characterized in that, include: The module includes a data collection module, a sub-skeleton analysis module, a 3D simulated molecular docking module, a Bayesian ridge regressor training module, a VAE generative model pre-training module, a reinforcement learning module, and a sampling evaluation module. The data collection module is used to collect drug molecule datasets, target-related ligand molecule datasets, and crystal structure data of target protein binding to ligands. The VAE generative model pre-training module is used to learn the basic syntax of SMILES molecular formulas and generate chemical molecule SMILES strings. The sub-skeleton analysis module uses ScaffoldGraph and a Bayesian theory-based sub-skeleton analysis method to perform sub-skeleton analysis on target-related ligand molecule datasets to obtain information on one or more sub-skeletons that contribute most to the activity. The 3D simulated molecular docking module performs 3D simulated docking on molecular samples in the target-related ligand molecule dataset and the crystal structure data of the target protein and ligand binding, to obtain the binding conformation and docking score containing molecular 3D information, as well as a dataset with binding energy. The Bayesian Ridge Regressor Training Module trains the Bayesian Ridge Regressor to fit the 3D docking results of molecules by using the molecular fingerprinting method on molecular samples with docking scores, thereby obtaining a regressor that predicts the docking scores of molecules. The reinforcement learning module optimizes the VAE generative model pre-training module by combining the Bayesian Ridge Regressor training module with a multi-objective optimization reward function formed by combining soft sub-skeleton constraints. The sampling evaluation module is used to test the VAE generation model generated by the reinforcement learning module. If it does not meet the requirements for drug design, the soft sub-scaffold constraints will be adjusted or the Bayesian ridge regressor training module will be adjusted.

2. The artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system according to claim 1, characterized in that, The VAE generative model pre-training module is used to learn the basic syntax of SMILES molecular formulas and generate chemical molecule SMILES strings; including: The AE model was trained using a combination of reconstruction loss and KL divergence loss; Each SMILES string is converted into a vector using embedding, and these vectors serve as input to the VAE model. The VAE model's encoder layer learns how to convert the input vector into a 50-dimensional continuous latent space vector, and the VAE model's decoder layer learns how to decode the original SMILES string from this 50-dimensional continuous latent space vector.

3. The artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system according to claim 2, characterized in that, When training a VAE model, the input includes not only the molecular formula of SMILES, but also the basic information of the molecules.

4. The artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system according to claim 1, characterized in that, When performing sub-scaffold analysis on the target-related ligand molecule dataset, other sub-scaffold structures can also be added manually.

5. The artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system according to claim 1, characterized in that, The 3D simulated molecular docking module performs 3D simulated docking on molecular samples from target-related ligand molecule datasets and crystal structure data of target protein-ligand binding, including: The dataset with binding energy obtained from the 3D simulated molecular docking module is input. First, the ECFP molecular fingerprint of each sample in 2048 dimensions is calculated using rdkit. Then, the molecular fingerprint is used as the feature matrix and the binding energy is used as the target vector to fit the model using a Bayesian ridge regressor. The model loss function is the mean squared error plus the L2 regularization term determined by the Bayesian method.

6. The artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system according to claim 1, characterized in that, The reinforcement learning module optimizes the VAE generative model pre-training module by combining the Bayesian ridge regressor training module with a multi-objective optimization reward function formed by soft sub-skeleton constraints; including: In each round of reinforcement learning, 500 latent space vectors of length 50 are sampled from the model, and these vectors are converted into 500 SMILES molecular expressions by the decoder. Next, calculate the efficiency and reward value of these molecules; Then, these reward values ​​are used to optimize the VAE generative model's sampling strategy in the latent space and the decoder layer weights.

7. The artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system according to claim 6, characterized in that, The multi-objective optimization reward function consists of multiple objectives and different weight values, including soft sub-skeleton constraints and a Bayesian ridge regressor.

8. The artificial intelligence-based molecular characterization and soft sub-scaffold-constrained drug discovery system according to claim 6, characterized in that, When optimizing the pre-training module of the VAE generative model, gradient descent is used, and the learning rate decay strategy is set to decrease by 10% every 50 rounds.

Images