Drug molecule generation optimization design system based on deep learning

Abstract

The invention discloses a drug molecule generation optimization design system based on deep learning, and the system comprises an input module which is used for receiving molecular structure information inputted by a user, analyzing an SMILES character string of a molecule, and converting the SMILES character string into a format which can be processed by the system; the processing module is connected with the input module and used for processing the SMILES character string according to a deep learning model of the Transform architecture and generating a new molecule; the optimization strategy module is connected with the processing module and used for selecting and executing a similarity-based optimization strategy or a multi-parameter-based optimization strategy and generating an optimized molecular structure; and the output module is connected with the optimization strategy module and is used for outputting the optimized new molecules. According to the method, the drug properties of molecules and the efficiency and effect of drug molecule generation optimization are remarkably improved, new molecules with excellent drug properties can be generated through flexible adjustment and optimization according to requirements, and higher optimization efficiency and better drug discovery effect are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of drug discovery, and in particular relates to a drug molecule generation optimization design system based on deep learning. Background Art

[0002] In the field of drug discovery, traditional molecular optimization methods mainly rely on experimental screening and rule-based algorithms. These methods usually require a lot of time and resources, and the optimization effect is limited. The experimental screening process is cumbersome, long, and costly, and it is difficult to cover a wide range of chemical spaces. Although rule-based algorithms can guide molecular design to a certain extent, they are difficult to cope with complex and changeable molecular structures and property requirements due to fixed rules and lack of flexibility. Therefore, traditional methods have significant deficiencies in both efficiency and effectiveness.

[0003] In response to the above problems, in recent years, some patents have attempted to use deep learning technology to improve the molecular optimization process. For example, patent CN119170151A discloses a drug molecule generation optimization method based on deep reinforcement learning. This method solves the generation optimization problem of drug molecules through a backbone network of LSTM and Transformer structures, and introduces a reinforcement learning dynamic correction model to design drugs that effectively bind to the target and have excellent chemical properties. In addition, patent CN202411276915.5 proposes a drug molecule optimization design method based on deep reinforcement learning and skeleton constraints, which optimizes the molecular generation model through transfer learning and reinforcement learning to improve the accuracy and efficiency of drug discovery.

[0004] Although the above patents have improved the efficiency and effectiveness of molecular optimization to a certain extent, there are still many shortcomings, such as high model complexity, high training costs, and single optimization strategy. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a drug molecule generation optimization design system based on deep learning, comprising:

[0006] The input module is used to receive the molecular structure information input by the user, parse the SMILES string of the molecule, and convert it into a format that can be processed by the system;

[0007] A processing module, connected to the input module, for processing the SMILES string according to a deep learning model of a Transformer architecture and generating a new molecule;

[0008] An optimization strategy module, connected to the processing module, for selecting and executing a similarity-based optimization strategy or a multi-parameter-based optimization strategy to generate an optimized molecular structure;

[0009] The output module is connected to the optimization strategy module and is used to output the optimized new molecules.

[0010] Preferably, the input module comprises a data preprocessing unit for standardizing and encoding the input SMILES character string;

[0011] The input module supports a variety of molecular input formats, including but not limited to SMILES strings and molecular structure diagrams.

[0012] Preferably, the deep learning model of the Transformer architecture adopts a multi-head self-attention mechanism, which includes multiple self-attention heads processed in parallel, and each self-attention head focuses on a different information subspace.

[0013] Preferably, the deep learning model of the Transformer architecture includes encoder and decoder layers;

[0014] Wherein, the encoder is used to extract features from input molecules;

[0015] The decoder is used to generate an optimized molecular structure.

[0016] Preferably, the deep learning model includes a self-attention layer and a feed-forward neural network layer;

[0017] The self-attention layer is used to capture long-range dependencies in molecular structures;

[0018] The feedforward neural network layer is used for nonlinear transformation;

[0019] The deep learning model uses transfer learning technology during the training process to initialize some network parameters using a pre-trained molecular feature extraction model;

[0020] The deep learning model adopts data enhancement technology during the training process to increase the diversity of training data by randomly perturbing the input SMILES string.

[0021] Preferably, the optimization strategy module includes a similarity evaluation unit and a multi-parameter evaluation unit;

[0022] The similarity evaluation unit is used to calculate the similarity between the new molecule and the target molecule;

[0023] The multi-parameter evaluation unit is used to evaluate the drug property parameters of the new molecule;

[0024] The similarity evaluation unit and the multi-parameter evaluation unit can be switched dynamically to adapt to different optimization requirements.

[0025] Preferably, the similarity evaluation unit calculates the molecular similarity using Tanimoto coefficient or cosine similarity.

[0026] Preferably, the multi-parameter evaluation unit comprises a plurality of submodules, each submodule being used to evaluate different drug property parameters of the new molecule;

[0027] The multi-parameter evaluation unit uses a multi-objective optimization algorithm to balance the optimization effects between different drug property parameters;

[0028] The drug property parameters include solubility, toxicity, and biological activity.

[0029] Preferably, the output module is used to output the optimized molecular structure in the form of a SMILES string and provide prediction results of relevant properties of the optimized molecule;

[0030] The output module includes a result visualization unit and a result export unit;

[0031] The result visualization unit is used to display the optimized new molecule in a graphical manner;

[0032] The result export unit is used to export the optimized new molecule data into multiple formats;

[0033] The formats include CSV files and JSON files.

[0034] Preferably, the system further comprises a feedback adjustment unit, a data storage unit, a model updating unit, a user interface unit, a log recording unit, a security verification unit, and a cloud service interface unit;

[0035] The feedback adjustment unit is connected to the output module and is used to receive user feedback and adjust parameter settings of the optimization strategy module;

[0036] The data storage unit is used to store the input SMILES character string, the intermediate data in the optimization process and the optimized new molecule data;

[0037] The model updating unit is used to regularly update the parameters of the deep learning model;

[0038] The user interface unit is used to provide a user interaction interface to support users to input optimization task parameters and view optimization results;

[0039] The log recording unit is used to record key information and system status during the optimization process;

[0040] The security verification unit is used to verify the user's identity and authority;

[0041] The cloud service interface unit is used to communicate with cloud computing resources and support the processing and result return of remote optimization tasks.

[0042] Compared with the prior art, the present invention has the following advantages and technical effects:

[0043] The present invention proposes a molecular optimization system based on deep learning, which adopts the Transformer architecture, can efficiently process long sequence molecular data, support multiple optimization strategies, and significantly improve the drug properties of molecules. The system generates new molecules similar to the target molecule or with specific properties by inputting the SMILES string of the molecule, effectively solving the shortcomings of the existing technology, bringing higher optimization efficiency and better drug discovery effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0045] Figure 1 is a schematic diagram of the first system structure of an embodiment of the present invention;

[0046] Figure 2 Schematic diagram of the second system structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0047] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0048] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0049] Embodiment 1

[0050] like Figure 1-2 As shown, this embodiment provides a drug molecule generation optimization design system based on deep learning, including:

[0051] The input module is used to receive the molecular structure information input by the user, specifically receive and parse the SMILES (Simplified Molecular Input Linear Expression) string of the molecule, and convert it into a format that can be processed by the system.

[0052] The processing module is connected to the input module and is used to process the SMILES string and generate new molecules according to the deep learning model of the Transformer architecture; specifically, the input molecules are feature extracted and optimized through the deep learning model of the Transformer architecture.

[0053] The optimization strategy module is connected to the processing module and is used to select and execute a similarity-based optimization strategy or a multi-parameter-based optimization strategy, and generate an optimized molecular structure according to the selected optimization strategy.

[0054] The optimization strategy module of this embodiment provides two molecular optimization strategies:

[0055] Among them, the similarity-based optimization strategy uses a deep learning model to generate new molecules similar to the target molecules by comparing the structural similarity between the input molecules and the target molecules.

[0056] The multi-parameter optimization strategy comprehensively considers various drug property parameters (such as solubility, bioavailability, etc.) and uses a multi-objective optimization algorithm to generate new molecules with specific properties.

[0057] An output module, connected to the optimization strategy module, is used to output the optimized new molecules;

[0058] The output module of this embodiment outputs the optimized molecular structure, specifically outputs the optimized molecular structure in the form of a SMILES string, and provides prediction results of relevant properties of the optimized molecule for user reference.

[0059] To further optimize the solution, the input module supports multiple molecular input formats, including but not limited to SMILES strings, molecular structure diagrams, etc.

[0060] Further optimizing the scheme, the deep learning model based on the Transformer architecture in this embodiment is used as the core processing unit of the processing module, responsible for molecular feature extraction and optimization. The Transformer architecture is adopted, and the self-attention mechanism is used to perform deep learning on the molecular structure. Based on the self-attention mechanism, the long-distance dependency in the molecular structure can be effectively captured, and the accuracy of molecular feature extraction can be improved.

[0061] The deep learning model of the Transformer architecture adopts a multi-head self-attention mechanism, which includes multiple self-attention heads processed in parallel, and each self-attention head focuses on a different information subspace.

[0062] The deep learning model of the Transformer architecture contains multiple encoder and decoder layers, where the encoder is used to extract features from the input molecule and the decoder is used to generate the optimized molecular structure.

[0063] The deep learning model includes multiple self-attention layers and feed-forward neural network layers. The self-attention layers are used to capture long-distance dependencies in molecular structures, and the feed-forward neural network layers are used for nonlinear transformations.

[0064] The deep learning model uses transfer learning technology during the training process and uses a pre-trained molecular feature extraction model to initialize some network parameters.

[0065] The deep learning model uses data augmentation technology during the training process to increase the diversity of training data by randomly perturbing the input SMILES string.

[0066] The input module further includes a data preprocessing unit for standardizing and encoding the input SMILES string.

[0067] To further optimize the scheme, the optimization strategy module includes a similarity evaluation unit and a multi-parameter evaluation unit;

[0068] The similarity assessment unit is used to calculate the similarity between the new molecule and the target molecule;

[0069] The Multi-Parameter Evaluation Unit is used to evaluate multiple pharmaceutical property parameters of new molecules.

[0070] The similarity evaluation unit and the multi-parameter evaluation unit in the optimization strategy module can be switched dynamically to adapt to different optimization requirements.

[0071] The similarity evaluation unit uses the Tanimoto coefficient or cosine similarity to calculate the molecular similarity.

[0072] The multi-parameter evaluation unit adopts a multi-objective optimization algorithm to balance the optimization effects among different drug property parameters.

[0073] The multi-parameter evaluation unit includes multiple sub-modules, each of which is used to evaluate different drug property parameters of new molecules, such as solubility, toxicity, biological activity, etc.

[0074] The output module includes a result visualization unit, which is used to display the optimized new molecules in a graphical manner.

[0075] The output module also includes a result export unit for exporting the optimized new molecule data into multiple formats, such as CSV files, JSON files, etc.

[0076] The system supports a distributed computing architecture and can process optimization tasks of multiple molecules in parallel.

[0077] The system also includes a feedback adjustment unit connected to the output module for receiving user feedback and adjusting parameter settings of the optimization strategy module.

[0078] The system also includes a data storage unit for storing the input SMILES character string, intermediate data in the optimization process and optimized new molecule data.

[0079] The system also includes a model updating unit for periodically updating the parameters of the deep learning model to maintain the optimized performance of the model.

[0080] The system also includes a user interface unit for providing a user interaction interface to support users in inputting optimization task parameters and viewing optimization results.

[0081] The system also includes a log recording unit for recording key information and system status during the optimization process to facilitate subsequent analysis and debugging.

[0082] The system also includes a security verification unit to verify user identity and permissions to ensure data security and stable system operation.

[0083] The system also includes a cloud service interface unit for communicating with cloud computing resources to support the processing and result return of remote optimization tasks.

[0084] The system of this embodiment uses the deep learning model of the Transformer architecture to achieve molecular optimization, and generates new molecules similar to the target molecule or with specific properties by inputting the SMILES string of the molecule. Through the above technical solution, the present invention can effectively optimize the molecule and improve its drug properties, which has important application value.

[0085] The drug molecule generation optimization design system of the present invention realizes efficient, flexible and accurate drug molecule optimization by combining advanced deep learning architecture (such as Transformer) and multiple optimization strategies, including:

[0086] 1. Efficiently process long sequence molecular data;

[0087] The present invention adopts a deep learning model based on the Transformer architecture, the core of which is the multi-head self-attention mechanism (Multi-Head Self-Attention). This mechanism can process multiple information subspaces in parallel and effectively capture long-distance dependencies in molecular structures. Compared with traditional recurrent neural networks (RNN) or convolutional neural networks (CNN), the Transformer architecture has significant advantages in processing long sequence molecular data (such as SMILES strings), and can significantly improve the efficiency and accuracy of feature extraction. For example, when processing complex polycyclic compounds, Transformer can more accurately identify the interactions between chemical bonds and functional groups within the molecule, thereby generating new molecular structures that are more in line with chemical laws.

[0088] 2. Flexible optimization strategy selection;

[0089] This system provides similarity-based optimization strategies and multi-parameter-based optimization strategies, and users can switch dynamically according to different needs. The similarity-based optimization strategy can quickly generate new molecules with similar structures to the target molecules by calculating the similarity between the new molecules and the target molecules (such as Tanimoto coefficient or cosine similarity), which is suitable for rapid screening and optimization of lead compounds. The multi-parameter-based optimization strategy comprehensively considers various drug property parameters such as solubility, toxicity, and biological activity, and balances the optimization effects between different parameters through a multi-objective optimization algorithm, so as to generate new molecules with specific properties. This flexible selection of optimization strategies enables the system to adapt to a variety of application scenarios from early drug discovery to late-stage drug optimization.

[0090] 3. Improve the pharmaceutical properties of drug molecules;

[0091] Through the multi-parameter evaluation unit in the optimization strategy module, the system is able to conduct a comprehensive evaluation and optimization of the drug properties of the generated new molecules. For example, during the optimization process, the system can simultaneously consider multiple parameters such as solubility, toxicity, and bioactivity, and generate drug molecules with both high bioactivity and low toxicity through multi-objective optimization algorithms (such as Pareto optimization). This comprehensive optimization strategy significantly improves the overall performance of drug molecules, reduces the failure rate of subsequent experimental screening, and increases the success rate of drug discovery.

[0092] 4. Application of data enhancement and transfer learning technology;

[0093] During the model training process, this system uses data augmentation technology and transfer learning technology. Data augmentation increases the diversity of training data and improves the generalization ability of the model by randomly perturbing the input SMILES string (such as randomly inserting, deleting or replacing characters). Transfer learning uses a pre-trained molecular feature extraction model to initialize some network parameters, allowing the model to converge faster on new tasks, reducing training time and computing resource consumption. This combination of technologies not only improves the training efficiency of the model, but also significantly improves the performance of the model on different molecular data sets.

[0094] 5. User-friendly and highly scalable;

[0095] This system provides a user-friendly interactive interface and supports a variety of molecular input formats (such as SMILES strings, molecular structure diagrams) and result output formats (such as CSV, JSON files). Users can easily input optimization task parameters through the interface and view the optimization results in real time. In addition, the system also supports distributed computing architecture and cloud service interfaces, which can process optimization tasks of multiple molecules in parallel and support the rapid optimization of large-scale drug molecules. This highly scalable design enables the system to meet different needs from laboratory research to industrial-level drug discovery.

[0096] 6. Real-time feedback and dynamic model update;

[0097] The system receives user feedback through the feedback adjustment unit and dynamically adjusts the parameter settings of the optimization strategy module based on the feedback. This real-time feedback mechanism can continuously optimize the performance of the model according to the actual needs of the user, ensuring that the generated molecular structure better meets the user's expectations. At the same time, the system is also equipped with a model update unit that can regularly update the parameters of the deep learning model to maintain the optimization performance of the model. This dynamic update mechanism enables the system to adapt to the ever-changing chemical data and optimization needs, and always maintain efficient and accurate optimization capabilities.

[0098] 7. Security and traceability;

[0099] The system is equipped with a security verification unit to verify user identity and permissions, ensuring data security and stable system operation. At the same time, the logging unit can record key information and system status during the optimization process to facilitate subsequent analysis and debugging. This security and traceability design makes the system more reliable when processing sensitive drug data, and also provides researchers with detailed experimental records to facilitate tracing every step of the optimization process.

[0100] In summary, the present invention significantly improves the efficiency and effect of drug molecule generation optimization by adopting advanced means such as the deep learning model of the Transformer architecture, multi-head self-attention mechanism, multi-parameter optimization strategy, data enhancement and transfer learning technology. The system can not only efficiently process complex molecular data, but also flexibly adjust the optimization strategy according to user needs to generate new molecules with excellent drug properties. In addition, the user-friendliness, scalability and safety design of the system enable it to be widely used in various stages of drug discovery, providing strong technical support for new drug research and development.

[0101] Embodiment 2: Figure 1 As shown, similarity-based molecular optimization;

[0102] Input molecule SMILES string: The user inputs the SMILES string of the target molecule, such as "CCO", through the input module.

[0103] The processing module standardizes the input SMILES string to ensure that it complies with chemical rules and encodes it into sequence data suitable for processing by the Transformer module.

[0104] The deep learning model based on the Transformer architecture in the processing module consists of multiple self-attention layers and feedforward neural network layers. The input coding sequence captures the internal dependencies of the molecule through the self-attention mechanism, and extracts and transforms features through the feedforward neural network to generate a new molecular structure code.

[0105] The generated molecular structure is then encoded and decoded into a SMILES string and chemical validity is verified to ensure that the generated molecular structure complies with chemical rules.

[0106] Output optimized molecules: Output the optimized molecule SMILES string to the user through the output module, such as "CCN".

[0107] Furthermore, in the deep learning model based on the Transformer architecture, the self-attention mechanism captures the complex relationships within the molecule by weighting the correlation between each element of the input sequence and all other elements. The feedforward neural network further processes the features output by the self-attention mechanism to generate a new molecular structure code.

[0108] Example 3: Molecular optimization based on multiple parameters;

[0109] like Figure 2 As shown, the system adds a multi-parameter optimization module on the basis of the second embodiment, which is used to perform molecular optimization according to multiple parameters specified by the user (such as solubility, toxicity, etc.).

[0110] The specific implementation steps include:

[0111] Input molecule SMILES string and optimization parameters: The user inputs the SMILES string and optimization parameters of the target molecule through the input module, such as "CCO, solubility>0.5, toxicity<0.1".

[0112] The processing module standardizes and encodes the input SMILES string and optimization parameters. The deep learning model based on the Transformer architecture receives the preprocessed data and generates preliminary molecular structure encoding through the self-attention mechanism and feedforward neural network.

[0113] The multi-parameter optimization module adjusts the initially generated molecular structure code according to the optimization parameters specified by the user to ensure that the generated new molecule meets the multi-parameter requirements specified by the user. The adjusted molecular structure code is then decoded into a SMILES string and chemical validity is verified.

[0114] The output module outputs the optimized molecular SMILES string to the user, such as "CCN".

[0115] Furthermore, in the multi-parameter optimization module, the system adjusts the molecular structure encoding according to the parameter requirements specified by the user through an optimization algorithm (such as gradient descent) to ensure that the generated new molecules meet all parameter requirements.

[0116] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A drug molecule generation optimization design system based on deep learning, characterized in that: include: The input module is used to receive the molecular structure information input by the user, parse the SMILES string of the molecule, and convert it into a format that can be processed by the system; A processing module, connected to the input module, for processing the SMILES string according to a deep learning model of a Transformer architecture and generating a new molecule; An optimization strategy module, connected to the processing module, for selecting and executing a similarity-based optimization strategy or a multi-parameter-based optimization strategy to generate an optimized molecular structure; The output module is connected to the optimization strategy module and is used to output the optimized new molecules.

2. The system according to claim 1, characterized in that The input module includes a data preprocessing unit for standardizing and encoding the input SMILES character string; The input module supports a variety of molecular input formats, including but not limited to SMILES strings and molecular structure diagrams.

3. The system according to claim 1, characterized in that The deep learning model of the Transformer architecture adopts a multi-head self-attention mechanism, which includes multiple self-attention heads processed in parallel, and each self-attention head focuses on a different information subspace.

4. The system according to claim 1, characterized in that The deep learning model of the Transformer architecture includes encoder and decoder layers; Wherein, the encoder is used to extract features from input molecules; The decoder is used to generate an optimized molecular structure.

5. The system according to claim 1, characterized in that The deep learning model includes a self-attention layer and a feed-forward neural network layer; The self-attention layer is used to capture long-range dependencies in molecular structures; The feedforward neural network layer is used for nonlinear transformation; The deep learning model uses transfer learning technology during the training process to initialize some network parameters using a pre-trained molecular feature extraction model; The deep learning model adopts data enhancement technology during the training process to increase the diversity of training data by randomly perturbing the input SMILES string.

6. The system according to claim 1, characterized in that The optimization strategy module includes a similarity evaluation unit and a multi-parameter evaluation unit; The similarity evaluation unit is used to calculate the similarity between the new molecule and the target molecule; The multi-parameter evaluation unit is used to evaluate the drug property parameters of the new molecule; The similarity evaluation unit and the multi-parameter evaluation unit can be switched dynamically to adapt to different optimization requirements.

7. The system according to claim 6, characterized in that The similarity evaluation unit calculates the molecular similarity using the Tanimoto coefficient or the cosine similarity.

8. The system according to claim 6, characterized in that The multi-parameter evaluation unit includes a plurality of submodules, each of which is used to evaluate different drug property parameters of the new molecule; The multi-parameter evaluation unit uses a multi-objective optimization algorithm to balance the optimization effects between different drug property parameters; The drug property parameters include solubility, toxicity, and biological activity.

9. The system according to claim 1, characterized in that The output module is used to output the optimized molecular structure in the form of a SMILES string and provide prediction results of relevant properties of the optimized molecule; The output module includes a result visualization unit and a result export unit; The result visualization unit is used to display the optimized new molecule in a graphical manner; The result export unit is used to export the optimized new molecule data into multiple formats; The formats include CSV files and JSON files.

10. The system according to claim 1, characterized in that The system also includes a feedback adjustment unit, a data storage unit, a model update unit, a user interface unit, a log recording unit, a security verification unit, and a cloud service interface unit; The feedback adjustment unit is connected to the output module and is used to receive user feedback and adjust parameter settings of the optimization strategy module; The data storage unit is used to store the input SMILES character string, the intermediate data in the optimization process and the optimized new molecule data; The model updating unit is used to regularly update the parameters of the deep learning model; The user interface unit is used to provide a user interaction interface to support users to input optimization task parameters and view optimization results; The log recording unit is used to record key information and system status during the optimization process; The security verification unit is used to verify the user identity and authority; The cloud service interface unit is used to communicate with cloud computing resources and support the processing and result return of remote optimization tasks.

Images