Traditional Chinese medicine recommendation device and method based on multi-scale molecular graph embedding and variational auto-encoder

Abstract

The invention discloses a traditional Chinese medicine recommendation device and method based on multi-scale molecular graph embedding and a variational auto-encoder, and the device comprises a data collection module which is used for collecting the molecular composition information and macroscopic function information of herbal medicines; the feature representation module is used for processing the molecular composition information and the macroscopic function information through the multi-scale molecular graph embedding module to generate feature representation of the herbal medicine; the missing data supplement module is used for generating supplement information of missing herbal medicine molecular structures through a variational auto-encoder model; and the recommendation module is used for carrying out herbal medicine recommendation according to the feature representation and the supplementary information. Molecular composition information and macroscopic function information of the herbal medicine are processed through the multi-scale molecular graph embedding module, feature representation of the herbal medicine is generated, supplementary information lacking herbal medicine molecular structures is generated through the variational auto-encoder model, a new technical means is provided for modernized research of traditional Chinese medicine, and the method is suitable for popularization and application. And the method has obvious application value and market prospect.

Description

Technical Field

[0001] The present invention belongs to the field of modern Chinese medicine research, and particularly relates to a Chinese medicine recommendation device and method based on multi-scale molecular graph embedding and variational autoencoder. Background Art

[0002] In the field of Chinese medicine research, with the continuous development of modern science and technology, how to efficiently and accurately represent the characteristics of Chinese medicine has become an important research direction. Traditional Chinese medicine research methods mainly rely on qualitative descriptions of herbal ingredients and simple statistical analyses. These methods are usually limited to describing the macroscopic functions of herbs and neglect the detailed information at the molecular level. Due to the lack of in-depth analysis of the molecular structures of herbs, traditional methods are difficult to provide feature representations with biological significance and cannot meet the needs of modern Chinese medicine research.

[0003] In recent years, with the development of artificial intelligence and machine learning technologies, some new methods have been introduced into Chinese medicine research. For example, Patent CN202310548996.9 proposes an intelligent Chinese medicine prescription recommendation system. This system uses a data input module, a data processing module, a Chinese medicine prescription generation module, and an output module to recommend prescriptions using the feature vectors of symptoms and herbs. However, this method is mainly based on simple feature vectors of symptoms and herbs and does not fully consider the complexity of the molecular structures of herbs, resulting in an incomplete feature representation.

[0004] Another patent CN202410930699.5 proposes a drug-target interaction prediction method based on graph attention network and multi-scale feature fusion. This method can better capture the structural information of drug molecules through multi-scale feature fusion. However, this method performs poorly in dealing with missing molecular structure information, affecting the accuracy of the overall analysis.

[0005] Although the above patents have improved the effect of Chinese medicine feature representation to a certain extent, there are still some deficiencies. For example, single-scale feature representation is difficult to comprehensively reflect the complex molecular structures of herbs, and existing methods lack an effective supplementary mechanism for dealing with missing molecular information. Summary of the Invention

[0006] The present invention proposes a Chinese medicine recommendation device and method based on multi-scale molecular graph embedding and variational autoencoder to solve the problems existing in the above prior art.

[0007] To achieve the above object, the present invention provides a Chinese medicine recommendation device based on multi-scale molecular graph embedding and variational autoencoder, including:

[0008] A data collection module for collecting the molecular composition information and macroscopic function information of herbs;

[0009] A feature representation module, configured to process the molecular composition information and macroscopic function information through a multi-scale molecular graph embedding module to generate a feature representation of the herbal medicine;

[0010] A missing data supplementation module, configured to generate supplementary information for the missing herbal medicine molecular structure through a variational autoencoder model;

[0011] A recommendation module, configured to perform herbal medicine recommendation based on the feature representation and supplementary information.

[0012] Preferably, the data collection module includes:

[0013] A data acquisition unit, configured to acquire the molecular composition information and macroscopic function information of the herbal medicine;

[0014] A preprocessing unit, configured to perform data cleaning and standardization processing on the molecular composition information and macroscopic function information;

[0015] A data storage unit, configured to store the preprocessed data.

[0016] Preferably, the feature representation module includes:

[0017] A first conversion unit, configured to convert the molecular composition information into a SMILES representation;

[0018] A second conversion unit, configured to convert the SMILES representation into a high-dimensional molecular embedding vector through a pre-trained molecular embedding model;

[0019] A feature generation unit, configured to combine the high-dimensional molecular embedding vector and the macroscopic function information of the herbal medicine to generate a feature representation with biological significance.

[0020] Preferably, the pre-trained molecular embedding model is a model based on a graph convolutional neural network.

[0021] Preferably, the missing data supplementation module includes:

[0022] A model construction unit, configured to construct a variational autoencoder model, where the variational autoencoder model includes an encoder part and a decoder part. The encoder part is configured to encode the input molecular structure data into a latent variable, and the decoder part is configured to decode the latent variable into a reconstructed molecular structure data;

[0023] A training unit, configured to train the variational autoencoder model using known herbal medicine molecular structure data;

[0024] A supplementation unit, configured to generate supplementary information for the missing herbal medicine molecular structure through the trained variational autoencoder model.

[0025] Preferably, the training unit includes:

[0026] A calculation unit for calculating the reconstruction error between the known molecular structure data and the reconstructed data;

[0027] An optimization unit for optimizing the parameters of the variational autoencoder model by minimizing the reconstruction error.

[0028] The present invention also provides a multi-scale molecular graph embedding method, including the following steps:

[0029] Obtain the molecular composition information of the herbal medicine and convert it into a SMILES representation;

[0030] Convert the SMILES representation into a high-dimensional molecular embedding vector through a pre-trained molecular embedding model;

[0031] Combine the high-dimensional molecular embedding vector with the macroscopic functional information of the herbal medicine to generate a biologically meaningful feature representation.

[0032] Preferably, the pre-trained molecular embedding model is a model based on a graph convolutional neural network;

[0033] The macroscopic functional information includes the medicinal properties of the herbal medicine and clinical trial data.

[0034] Preferably, the generation of the biologically meaningful feature representation includes:

[0035] Perform dimensionality reduction processing on the high-dimensional molecular embedding vector;

[0036] Fuse the vector after dimensionality reduction with the macroscopic functional information to generate a comprehensive feature representation.

[0037] The present invention also provides a traditional Chinese medicine recommendation method based on the multi-scale molecular graph embedding method, including the following steps:

[0038] Construct a variational autoencoder model and train the variational autoencoder model using known herbal medicine molecular structure data;

[0039] Use the trained variational autoencoder model to generate supplementary information for missing herbal medicine molecular structures;

[0040] Perform herbal medicine recommendation according to the feature representation and the supplementary information;

[0041] Performing herbal medicine recommendation includes: calculating the similarity between different herbal medicine feature representations;

[0042] Sort the herbal medicines according to the similarity to generate a recommendation list.

[0043] Compared with the prior art, the present invention has the following advantages and technical effects:

[0044] The present invention discloses a traditional Chinese medicine recommendation device and method based on multi-scale molecular graph embedding and variational autoencoder, including: a data collection module for collecting molecular composition information and macroscopic function information of herbs; a feature representation module for processing the molecular composition information and macroscopic function information through a multi-scale molecular graph embedding module to generate a feature representation of the herbs; a missing data supplementation module for generating supplementary information of the missing molecular structure of the herbs through a variational autoencoder model; and a recommendation module for recommending herbs according to the feature representation and the supplementary information. The present invention processes the molecular composition information and macroscopic function information of herbs through a multi-scale molecular graph embedding module to generate a feature representation of the herbs, and generates supplementary information of the missing molecular structure of the herbs through a variational autoencoder model, providing a new technical means for the modern research of traditional Chinese medicine, and having significant application value and market prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0046] Figure 1 It is a schematic diagram of the device according to an embodiment of the present invention;

[0047] Figure 2 It is a flowchart of the embedding method according to an embodiment of the present invention;

[0048] Figure 3 It is a flowchart of the supplementary information generation method according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments may be combined with each other. The following will refer to the drawings and combine the embodiments to detail this application.

[0050] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0051] The present invention deeply analyzes and characterizes the molecular structure and macroscopic function information of traditional Chinese medicine through advanced technical means. The system generates a biologically meaningful feature representation by digitally representing the molecular composition information of herbs, fusing features, and supplementing missing data, and realizes personalized herb recommendations based on this. This process mainly serves the research, development, and optimization of traditional Chinese medicine, providing new technical tools and methods for the modern research of traditional Chinese medicine. The present invention is not directly used for the treatment or diagnosis of diseases, but focuses on improving the scientific nature and efficiency of traditional Chinese medicine research, and promoting technological innovation and theoretical development in the field of traditional Chinese medicine.

[0052] Example 1

[0053] As Figure 1 shown, in this embodiment, a traditional Chinese medicine recommendation device based on multi-scale molecular graph embedding and variational autoencoder is provided, including:

[0054] A data collection module, which is used to collect the molecular composition information and macroscopic function information of herbs. The molecular composition information of herbs includes but is not limited to molecular formula, molecular structure, etc.; the macroscopic function information includes the medicinal properties of herbs and clinical experiment data;

[0055] A feature representation module, which is used to process the molecular composition information and macroscopic function information through a multi-scale molecular graph embedding module to generate a feature representation of herbs;

[0056] A missing data supplement module, which is used to generate supplementary information for missing herb molecular structures through a variational autoencoder model;

[0057] A recommendation module, which is used to make herb recommendations according to the feature representation and supplementary information. When making recommendations, algorithms such as collaborative filtering and deep learning are used to make personalized recommendations in combination with user needs and herb features.

[0058] Further, the data collection module includes:

[0059] A data acquisition unit, which is used to obtain the molecular composition information and macroscopic function information of herbs;

[0060] A preprocessing unit, which is used to perform data cleaning and standardization processing on the molecular composition information and macroscopic function information;

[0061] A data storage unit, which is used to store the preprocessed data.

[0062] Further, the feature representation module includes:

[0063] A first conversion unit, which is used to convert the molecular composition information into SMILES representation;

[0064] A second conversion unit, configured to convert the SMILES representation into a high-dimensional molecular embedding vector through a pre-trained molecular embedding model;

[0065] A feature generation unit, configured to combine the high-dimensional molecular embedding vector with the macroscopic functional information of the herbal medicine to generate a biologically meaningful feature representation.

[0066] Further, the pre-trained molecular embedding model is a model based on a graph convolutional neural network.

[0067] Further, the missing data supplementation module includes:

[0068] A model construction unit, configured to construct a variational autoencoder model, where the variational autoencoder model includes an encoder part and a decoder part. The encoder part is configured to encode the input molecular structure data into a latent variable, and the decoder part is configured to decode the latent variable into reconstructed molecular structure data;

[0069] A training unit, configured to train the variational autoencoder model using known herbal medicine molecular structure data;

[0070] A supplementation unit, configured to generate supplementary information of missing herbal medicine molecular structures through the trained variational autoencoder model.

[0071] Further, the training unit includes:

[0072] A calculation unit, configured to calculate the reconstruction error between the known molecular structure data and the reconstructed data;

[0073] An optimization unit, configured to optimize the parameters of the variational autoencoder model by minimizing the reconstruction error.

[0074] Embodiment 2

[0075] As Figure 2 shown, a multi-scale molecular graph embedding method is provided in this embodiment, including the following steps:

[0076] Obtain the molecular composition information of the herbal medicine and convert it into a SMILES representation;

[0077] Convert the SMILES representation into a high-dimensional molecular embedding vector through a pre-trained molecular embedding model;

[0078] Combine the high-dimensional molecular embedding vector with the macroscopic functional information of the herbal medicine to generate a biologically meaningful feature representation.

[0079] Specifically, it includes the following steps:

[0080] 1 Data collection and preprocessing:

[0081] First, collect the molecular composition information of various herbs. Taking common Chinese herbal medicines such as ginseng, astragalus, and angelica as examples, obtain their molecular composition data through chemical analysis methods. Convert these molecular composition data into SMILES (Simplified Molecular Input Line Entry System) representation. For example, the SMILES representation of the main component ginsenoside Rg1 in ginseng is "COC1=CC=CC2=CC(C3=CC=CC4=CC(C5=CC=CC6=CC(C7=CC=CC8=CC(C9=CC=CC1)=O)=O)=O)=O)=O)=O".

[0082] 2 Pretrained molecular embedding model:

[0083] Use a pretrained molecular embedding model, such as Graph Convolutional Network (GCN) or Graph Attention Network (GAT), to convert the molecular structure represented by SMILES into a high-dimensional molecular embedding vector. The specific steps are as follows:

[0084] 1. Molecular graph construction: Parse the SMILES string into a molecular graph, where nodes represent atoms and edges represent chemical bonds.

[0085] 2. Feature extraction: Extract features for each node and edge, such as atom type, chemical bond type, etc.

[0086] 3. Graph embedding: Use the GCN or GAT model to embed the molecular graph and generate a high-dimensional molecular embedding vector.

[0087] 3 Multi-scale feature fusion:

[0088] Fuse the generated molecular embedding vector with the macroscopic functional features of the herbs. The macroscopic functional features include the medicinal effects and properties of the herbs, which can be obtained through literature research and experimental data. The fusion method uses weighted average or attention mechanism. The specific steps are as follows:

[0089] 1. Feature standardization: Standardize the molecular embedding vector and macroscopic functional features.

[0090] 2. Weight assignment: Assign weights to different features according to the specific application scenarios of the herbs.

[0091] 3. Feature fusion: Fuse the weighted features to generate the final multi-scale feature representation.

[0092] 4 Application examples:

[0093] Taking ginseng as an example, the multi-scale feature representations generated through the above steps can be used in a traditional Chinese medicine recommendation system. The system matches the most suitable combination of herbs according to the symptoms and constitution input by the user. Experiments show that the MMGE method significantly improves the accuracy and robustness of the recommendation system.

[0094] Furthermore, the pre-trained molecular embedding model is a model based on a graph convolutional neural network;

[0095] The macroscopic functional information includes the medicinal properties of herbs and clinical experimental data.

[0096] Furthermore, generating biologically meaningful feature representations includes:

[0097] Performing dimensionality reduction on the high-dimensional molecular embedding vectors;

[0098] Fusing the vectors after dimensionality reduction with the macroscopic functional information to generate a comprehensive feature representation.

[0099] Example Three

[0100] A traditional Chinese medicine recommendation method based on a multi-scale molecular graph embedding method, including the following steps:

[0101] Construct a variational autoencoder model and train the variational autoencoder model using known herbal molecular structure data;

[0102] Using the trained variational autoencoder model, generate supplementary information for missing herbal molecular structures;

[0103] Based on the feature representations and supplementary information, perform herbal recommendations;

[0104] Performing herbal recommendations includes: calculating the similarity between different herbal feature representations;

[0105] Sort the herbs according to the similarity to generate a recommendation list.

[0106] Furthermore, the process of generating supplementary information is as Figure 3 shown, specifically including:

[0107] 1 Dataset construction:

[0108] Collect the molecular structure information of herbs from public datasets to construct a dataset containing complete and missing molecular structures. For example, the SMILES representation of some herbs may be missing, such as only partial structural information for a certain component in Astragalus membranaceus.

[0109] 2 Variational autoencoder (VAE) model training:

[0110] Use the VAE model to train the complete molecular structure data to learn the latent representation of the molecular structure.

[0111] The specific steps are as follows:

[0112] 1. Encoder design: Design an encoder network to transform the molecular graph into a latent space vector.

[0113] 2. Decoder design: Design a decoder network to reconstruct the latent space vector into a molecular graph.

[0114] 3. Loss function: Adopt the reconstruction loss and KL divergence loss to optimize the VAE model.

[0115] 3. Supplementary missing molecular structure:

[0116] Use the trained VAE model to supplement the missing molecular structure. The specific steps are as follows:

[0117] 1. Input of partial structure: Input the structure information of the missing part into the encoder.

[0118] 2. Generation of latent vector: Generate a latent space vector.

[0119] 3. Structure reconstruction: Reconstruct the complete molecular structure through the decoder.

[0120] 4. Application example:

[0121] Taking Astragalus membranaceus as an example, the missing molecular structure information is supplemented through the VAE model to generate a complete SMILES representation. The supplemented molecular structure information is used in the MMGE method to further improve the integrity and accuracy of the traditional Chinese medicine feature representation.

[0122] In summary, the present invention significantly improves the accuracy and integrity of the traditional Chinese medicine feature representation through the multi-scale molecular graph embedding method and the method for supplementing the missing molecular structure features based on the variational autoencoder, providing a new technical means for the modern research of traditional Chinese medicine.

[0123] The above is only a preferred specific embodiment 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 those skilled in the art within the technical scope disclosed in the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A traditional Chinese medicine recommendation device based on multi-scale molecular graph embedding and variational autoencoder, characterized in that, It includes: A data collection module for collecting the molecular composition information and macroscopic function information of herbs; A feature representation module for processing the molecular composition information and macroscopic function information through a multi-scale molecular graph embedding module to generate a feature representation of herbs; A missing data supplementation module for generating supplementary information on missing herb molecular structures through a variational autoencoder model; A recommendation module for making herb recommendations based on the feature representation and supplementary information.

2. The device according to claim 1, characterized in that, The data collection module includes: A data acquisition unit for obtaining the molecular composition information and macroscopic function information of herbs; A preprocessing unit for performing data cleaning and standardization processing on the molecular composition information and macroscopic function information; A data storage unit for storing the preprocessed data.

3. The device according to claim 1, characterized in that, The feature representation module includes: A first conversion unit for converting the molecular composition information into a SMILES representation; A second conversion unit for converting the SMILES representation into a high-dimensional molecular embedding vector through a pre-trained molecular embedding model; A feature generation unit for combining the high-dimensional molecular embedding vector with the macroscopic function information of herbs to generate a feature representation with biological significance.

4. The device according to claim 3, characterized in that, The pre-trained molecular embedding model is a model based on a graph convolutional neural network.

5. The device according to claim 1, characterized in that, The missing data supplementation module includes: A model construction unit for constructing a variational autoencoder model, which includes an encoder part and a decoder part. The encoder part is used to encode the input molecular structure data into a latent variable, and the decoder part is used to decode the latent variable into reconstructed molecular structure data; A training unit for training the variational autoencoder model using known herb molecular structure data; A supplementation unit for generating supplementary information on missing herb molecular structures through the trained variational autoencoder model.

6. The device according to claim 5, characterized in that The training unit includes: A calculation unit for calculating the reconstruction error between the known molecular structure data and the reconstructed data; An optimization unit for optimizing the parameters of the variational autoencoder model by minimizing the reconstruction error.

7. A multi-scale molecular graph embedding method, characterized in that, It includes the following steps: Obtain the molecular composition information of herbs and convert it into a SMILES representation; Through a pre-trained molecular embedding model, convert the SMILES representation into a high-dimensional molecular embedding vector; Combine the high-dimensional molecular embedding vector with the macroscopic function information of herbs to generate a feature representation with biological significance.

8. The method according to claim 7, wherein The pre-trained molecular embedding model is a model based on a graph convolutional neural network; The macroscopic function information includes the properties of herbs and clinical trial data.

9. The method according to claim 7, characterized in that, The generation of the feature representation with biological significance includes: Performing dimensionality reduction processing on the high-dimensional molecular embedding vector; Fusing the dimensionality-reduced vector with the macroscopic function information to generate a comprehensive feature representation.

10. A traditional Chinese medicine recommendation method based on the method described in claim 7, characterized in that, It includes the following steps: Construct a variational autoencoder model and train the variational autoencoder model using known herb molecular structure data; Utilize the trained variational autoencoder model to generate supplementary information on missing herb molecular structures; Based on the feature representation and supplementary information, make herb recommendations; Performing herbal medicine recommendations includes: calculating the similarity between different herbal medicine feature representations; Sorting the herbal medicines according to the similarity to generate a recommendation list.

Images