Chinese herbal medicine indication discovery method and system based on hypergraph convolution

Abstract

The invention provides a Chinese herbal medicine indication discovery method and system based on hypergraph convolution. The method is mainly based on two hypergraph convolutions, and makes full use of biological knowledge features of effective components and target spots, high-dimensional Chinese herbal medicine-effective component association and target spot-indication association including molecular fingerprints of the effective components, sequence information of the target spots and high-order features in hypergraphs. The method is low in model complexity and low in power consumption; the association relationship between the Chinese herbal medicine and the indications in the network can be accurately predicted, and reliable technical support is provided for finding the indications of the Chinese herbal medicine.

Description

Technical Field

[0001] The present application belongs to the field of machine learning in computer science and bioinformatics in biology, and specifically relates to a method and system for discovering indications of Chinese herbal medicines based on hypergraph convolution. Background Art

[0002] With thousands of years of history, Chinese herbal medicine has been widely used to treat and prevent difficult and complicated diseases, and has accumulated valuable experience in medication. Unlike modern drugs that are composed of a single ingredient, Chinese herbal medicine contains multiple active ingredients and can therefore act on multiple molecular biological mechanisms within cells simultaneously. Because of this, Chinese herbal medicines are effective for different indications from a systemic perspective. However, the main difficulty in understanding the association between Chinese herbal medicines and indications lies in the multi-target and multi-component mechanisms, which demonstrates the inherent complexity of Chinese herbal medicines in treating indications. Therefore, it is necessary to consider multi-target and multi-component mechanisms to improve the performance of Chinese herbal medicine-indication association prediction, and this challenging task remains to be solved.

[0003] Although existing technologies have achieved certain research results in the discovery of indications for Chinese herbal medicines, most existing methods mainly focus on predicting the association between active ingredients and target proteins or between Chinese herbal medicines and target proteins. Most of these prediction models fail to incorporate multi-component and multi-target mechanisms. At the same time, most studies focus more on the association between Chinese herbal medicines and symptoms, and fail to study the mechanism of action of Chinese herbal medicines from the perspective of molecular association. Therefore, using the multi-component and multi-target mechanism of Chinese herbal medicines makes the prediction ability of the model more accurate, thereby solving the problem of discovering indications for Chinese herbal medicines, and can provide better solutions for drug discovery and clinical medication. Summary of the invention

[0004] The purpose of this application is to overcome the defects of the prior art that multiple targets and multiple component mechanisms are not considered.

[0005] In order to achieve the above objectives, this application proposes a method for discovering indications of Chinese herbal medicines based on hypergraph convolution, comprising:

[0006] S1: Collect data on the relationship between Chinese herbal medicines, active ingredients, target sites and indications verified by biological experiments; obtain the structural data of active ingredients and sequence information of target sites;

[0007] S2: According to the original data obtained in step S1, the fingerprint similarity of the active ingredients is calculated and the active ingredient similarity matrix is ​​generated, and the similarity of the target sequence is calculated and the target sequence similarity matrix is ​​generated;

[0008] S3: Use an autoencoder to reduce the dimension of the similarity matrix of the active ingredient and the target to obtain low-dimensional features of the active ingredient and the target;

[0009] S4: construct corresponding hypergraphs based on the collected Chinese herbal medicine-active ingredient association information and target-indication association information;

[0010] S5: using the low-dimensional features of the active ingredients and target points obtained in step S3 as the initial features of the nodes in the hypergraph, and using two hypergraph convolutional networks to learn the high-order structural information in the Chinese herbal medicine-active ingredients and target-indication hypergraphs, thereby obtaining the feature representation of Chinese herbal medicines and indications in the network;

[0011] S6: Using the acquired characteristic representations of the Chinese herbal medicine and indications as inputs to the scoring function to obtain scores for different Chinese herbal medicine-indication association pairs; judging whether there is an association between the Chinese herbal medicine and the indication based on the scores.

[0012] As an improvement of the above method, step S2 includes:

[0013] Step S2-1: Convert the SMILES data of the active ingredient into a molecular fingerprint. When quantifying the similarity between two molecular fingerprints, the Dice coefficient is used to measure the overlap between the substructures represented by the fingerprints. Assume that c i and c j The Dice coefficient between two effective components is calculated as:

[0014]

[0015] Among them, c i and c j Represents two Chinese herbal medicine ingredients; M i and M j c i and c j The molecular fingerprint of To express the similarity between pairs of effective components, where n c is the number of components;

[0016] Step S2-2: Obtain the sequence information of the target, calculate the similarity between the target sequences, and construct a similarity matrix, which is expressed as where n t is the number of target points.

[0017] As an improvement of the above method, step S4 includes:

[0018] An active ingredient-Chinese herbal medicine hypergraph with active ingredients as nodes and Chinese herbal medicines as hyperedges and a target target-indication hypergraph with target targets as nodes and indications as hyperedges are constructed respectively.

[0019] As an improvement of the above method, step S5 includes:

[0020] Step S5-1: using low-dimensional features of active ingredients and target points as initial vectors of nodes of the hypergraph of active ingredients-Chinese herbal medicine and the hypergraph of target points-indications;

[0021] Step S5-2: Learn the feature representation of hyperedges and mine potential associations through the hypergraph convolutional neural network, including:

[0022] Step S5-2-1: Hyperedge aggregation:

[0023] The hyperedge aggregation formula of the active ingredient-Chinese herbal medicine hypergraph is:

[0024]

[0025] in, is the indicator matrix A of the active ingredient-Chinese herbal medicine hypergraph ch The edge degree matrix of ch and M ch is the embedding matrix of active components and hyperedges at layer l; information from active component nodes to their hyperedges is transmitted through The multiplication operation is realized; the superscript T represents the matrix transpose;

[0026] The hyperedge aggregation formula of the target-indication hypergraph is:

[0027]

[0028] in, is the indicator matrix A of the target-indication hypergraph td The edge degree matrix of td and M td is the target node and hyperedge embedding matrix of the lth layer; the information from the target node to its hyperedge is transmitted through The multiplication operation is realized;

[0029] Step S5-2-2: Node aggregation:

[0030] The node aggregation process of the active ingredient-Chinese herbal medicine hypergraph is expressed as:

[0031] The conversion process from hyperedge to node is:

[0032]

[0033] in, A ch The node degree matrix of A is used to rescale the effective component embedding; ch With M chMultiply,aggregate information from hyperedges to active component nodes;

[0034] The complete hypergraph convolution is defined as follows:

[0035]

[0036] Among them, θ l is a trainable parameter; σ(·) is a nonlinear activation function; W C The encoding process of the automatic encoder in step S3:

[0037]

[0038] Among them, W enc is the weight matrix; b enc is the bias vector of the encoder; g(·) is the nonlinear activation function; For S C The i-th row vector of S C for the similarity between the active ingredients;

[0039] The node aggregation process of the target-indication hypergraph is expressed as:

[0040] The conversion process from hyperedge to node is:

[0041]

[0042] in, A td The node degree matrix of A is used to rescale the effective component embedding; td With M td Multiply, aggregate information from the hyperedge to the target node;

[0043] The complete hypergraph convolution is defined as follows:

[0044]

[0045] in, W T The encoding process of the automatic encoder in step S3:

[0046]

[0047] in, For S T The i-th row vector of T is the similarity between the target sites;

[0048] Step S5-2-3: Model training: using cross entropy as the loss function for model training.

[0049] As an improvement of the above method, step S6 includes:

[0050] Perform inner product operations on the characteristic representations of Chinese herbal medicines and indications in the network to obtain the scores between different associations

[0051]

[0052] Among them, E H and E D Represent the characteristics of Chinese herbal medicine and the characteristics of indications respectively; the superscript T indicates matrix transposition.

[0053] The present application also provides a Chinese herbal medicine indication discovery system based on hypergraph convolution, which is implemented based on the above method, and the system includes:

[0054] The data collection module is used to collect data on the relationship between Chinese herbal medicines, active ingredients, target sites and indications verified by biological experiments; obtain the structural data of active ingredients and the sequence information of target sites;

[0055] A data preprocessing module, used to calculate the fingerprint similarity of the active ingredients and generate an active ingredient similarity matrix, calculate the similarity of the target sequence and generate a target sequence similarity matrix based on the original data obtained by the data collection module;

[0056] The module for acquiring low-dimensional features of active ingredients and target points is used to reduce the dimension of the similarity matrix of active ingredients and target points by using an automatic encoder to acquire low-dimensional features of active ingredients and target points;

[0057] A hypergraph construction module is used to construct corresponding hypergraphs according to the collected Chinese herbal medicine-active ingredient association information and target point-indication association information;

[0058] The network high-order structural feature generation module is used to use the low-dimensional features of the active ingredients and target targets obtained by the low-dimensional feature acquisition module of the active ingredients and target targets as the initial features of the nodes in the hypergraph, and use two hypergraph convolutional networks to learn the high-order structural information in the Chinese herbal medicine-active ingredient and target target-indication hypergraphs, so as to obtain the feature representation of Chinese herbal medicine and indications in the network;

[0059] The potential Chinese herbal medicine-indication association prediction module is used to use the acquired feature representations of Chinese herbal medicines and indications as inputs to the scoring function to obtain scores for different Chinese herbal medicine-indication association pairs; and to determine whether there is an association between the Chinese herbal medicines and indications based on the scores.

[0060] Compared with the prior art, the advantages of this application are:

[0061] 1. The hypergraph convolution method proposed in the present invention uses the biological knowledge features of the nodes in the graph to initialize the nodes, which can solve the cold start problem and has both physical and practical significance. It ensures the distance relationship between the effective component nodes and the target target nodes in their own similarity space, and learns the characteristic representation of Chinese herbal medicines and indications by constructing the Chinese herbal medicine-effective ingredient hypergraph and the target target-indication hypergraph, thereby ensuring that Chinese herbal medicines with high similarity have similar therapeutic effects on a certain type of indication, which meets the practical significance of indication treatment;

[0062] 2. The present invention makes full use of the latent vector obtained by the autoencoder and uses it as the initial vector of the nodes in the network representation of Chinese herbal medicine and indications in the next step, so that the nodes retain attribute information and correlations with each other. When learning higher-dimensional features, the model is easy to train and more stable;

[0063] 3. The present invention makes full use of the autoencoder and hypergraph convolutional network to obtain the characteristic representation of Chinese herbal medicine and indications in the hypergraph. The autoencoder is used to reduce the dimension of the similarity matrix of the active ingredients and the target to obtain the initial vector of the node in the hypergraph, which is helpful to complete other downstream tasks, such as exploring the interaction between different types of nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Shown is a flowchart of the method for discovering indications of Chinese herbal medicines based on hypergraph convolution. DETAILED DESCRIPTION

[0065] The technical solution of the present application is described in detail below with reference to the accompanying drawings.

[0066] The purpose of the present invention is to provide a method and system for discovering Chinese herbal medicine indications based on hypergraph convolution. The method is mainly based on two hypergraph convolutions, making full use of the biological knowledge characteristics of active ingredients and target targets, as well as high-dimensional Chinese herbal medicine-active ingredient associations and target target-indication associations, including molecular fingerprints of active ingredients, sequence information of target targets, and high-order features in the hypergraph. The method of the present invention has low model complexity and low power consumption; it can accurately predict the association between Chinese herbal medicines and indications in the network, and provide reliable technical support for the discovery of drug indications for Chinese herbal medicines.

[0067] like Figure 1 As shown, the present invention provides a method for discovering indications of Chinese herbal medicines based on hypergraph convolution, comprising the following steps:

[0068] S1. Select and establish data sets: Based on published literature, collect data on the relationship between Chinese herbal medicines, active ingredients, target sites and indications verified by biological experiments; obtain the structural data of active ingredients based on the PubChem database; obtain the sequence information of target sites based on the Ensembl database;

[0069] S2. Data preprocessing: According to the raw data obtained in step S1, by applying statistical knowledge and informatics theory, the fingerprint similarity of the active ingredients is calculated and the active ingredient similarity matrix is ​​generated, and the similarity of the target sequence is calculated and the target sequence similarity matrix is ​​generated;

[0070] S3. Obtain low-dimensional features of active ingredients and target targets: Use an autoencoder to reduce the dimension of the similarity matrix of the active ingredients and target targets, and minimize the error between the reduced vector and the original vector.

[0071] S4. Constructing hypergraphs: constructing different hypergraphs based on the collected information on the association between Chinese herbal medicines and active ingredients and the association between target points and indications;

[0072] S5. Generate high-order structural features of the network: Use the low-dimensional vectors of active ingredients and target points obtained in S3 as the initial features of the nodes in the hypergraph, and use two hypergraph convolutional networks to learn the high-order structural information in the Chinese herbal medicine-active ingredient and target-indication hypergraphs, so as to obtain the feature representation of Chinese herbal medicine and indications in the network;

[0073] S6. Predict potential Chinese herbal medicine-indication associations: Use the acquired feature representations of Chinese herbal medicines and indications as inputs to the scoring function to obtain scores for different Chinese herbal medicine-indication association pairs, and then determine whether there is an association between the Chinese herbal medicines and indications based on the acquired scores.

[0074] The following is a detailed introduction to each step:

[0075] S1. Select and establish data set;

[0076] This example uses the latest benchmark dataset, collected from the TCM-suite database. Specifically, the TCM-suite database consists of Chinese herbal medicine-active ingredient associations, active ingredient-target protein associations, and target protein-indication associations, and is divided into two sub-databases: Holmes-Suite focuses on Chinese herbal medicine ingredient identification, while Watson-Suite facilitates pharmacological network analysis. A total of 1,009 Chinese herbal medicines, 1,193 active ingredients, 7,258 target proteins, 11,071 indications, 6,496 Chinese herbal medicine-active ingredient associations, and 44,170 target protein-indication associations were collected, which represent the relationships between Chinese herbal medicines, active ingredients, target proteins, and indications in the database. There may be 11,170,639 associations between 1,009 Chinese herbal medicines and 11,071 indications. After integrating the data of Chinese herbal medicine-active ingredient association, active ingredient-target protein association, and target protein-indication association, we found that there were 2,354,225 known associations between all Chinese herbal medicines and indications. Canonical SMILES information of active ingredients was collected from the PubChem database, and sequence information of target sites was collected from the Ensembl database;

[0077] S2. Data preprocessing;

[0078] The first step is to establish the molecular fingerprint similarity matrix of active ingredients. The SMILES data of Chinese herbal medicine ingredients were converted into molecular fingerprints using the RDKit tool. When quantifying the similarity between two Morgan fingerprints, the Dice coefficient was used to measure the degree of overlap between the substructures represented by the fingerprints. Assume that c i and c j are two Chinese herbal medicine ingredients, and the Dice coefficient between them is calculated as:

[0079]

[0080] Among them, M i and M j They are c i and c j Morgan fingerprint, so we can get a similarity matrix To express the similarity between pairs of effective components, where n c is the number of components.

[0081] The second step is to build a target sequence similarity matrix. In this example, the sequence information of the target is obtained through an open source database on the Internet, and the similarity between the target protein sequences is calculated using the PairwiseAligner module in the Biopython tool. The parameter settings used when learning the target protein similarity are the same as the default settings provided in the Biopython tool. Assume that n t is the number of target proteins, and finally a similarity matrix can be constructed, expressed as

[0082] S3. Obtain low-dimensional features of active ingredients and target points;

[0083] This step mainly reduces the dimension of the effective ingredient similarity matrix and target similarity matrix obtained by the data processing module. C and S T The dimension of S is high, so it is not appropriate to use it as the feature vector of the active ingredient and the target. Therefore, this example uses the autoencoder model to obtain the low-dimensional embedding of the active ingredient and the target. Specifically, the autoencoder is an unsupervised learning model based on a deep neural network, which consists of an encoder and a decoder. The autoencoder aims to convert S C and S T The row vector of is compressed into a low-dimensional embedding. For example, let For S C The i-th row vector of , the encoding process can be expressed as:

[0084]

[0085] Among them, W enc is the weight matrix, b enc is the bias vector of the encoder, and g(·) is a nonlinear activation function such as RELU or Tanh. With (2), we can transform the high-dimensional input vector Mapped to a low-dimensional latent representation W C .

[0086] The decoder in the autoencoder is responsible for C Reconstruct S C To do this, it performs the inverse transform of the encoder and tries to reconstruct S as accurately as possible. C The decoding process can be expressed as:

[0087]

[0088] Among them, W dec is the weight matrix, b dec is the bias vector, yes In order to approximate This example uses the mean square error (MSE) as the loss function, making and The error between them is minimized. The definition of MSE is as follows:

[0089]

[0090] During the training process of the autoencoder, the trainable parameters are updated using the back-propagation algorithm to minimize the reconstruction error, which is defined as the original component similarity vector according to (5). and their reconstructed counterparts The average loss between .

[0091]

[0092] In the above formula, n is the number of effective components contained in the data set. The autoencoder is based on S C and S T The training process aims to learn W C and W T , which are the embedding matrices of effective ingredients and target points respectively. By retaining the information consistent with the original feature space, W C and W T The embedding matrix in can capture the important features and patterns of active ingredients and target points, which can be used as valuable input features to initialize the nodes in the hypergraph.

[0093] S4. Construct a hypergraph;

[0094] This step mainly constructs a hypergraph with active ingredients as nodes and Chinese herbal medicines as hyperedges and a hypergraph with target targets as nodes and indications as hyperedges through the Chinese herbal medicine-active ingredient association and target target-indication association information obtained in step S1. The hypergraph is represented as in:

[0095] V={v i} represents the node set in the hypergraph, with a length of m;

[0096] ε={e j} represents a hyperedge set with a length of n;

[0097] A∈{0,1} m×n is an indicator matrix used to indicate whether a node and a hyperedge exist in the hypergraph. Each element a(v i ,e j ) represents node v i Is there a hyperedge e j middle.

[0098] If a(v i ,e j )=1, it means v i Yes j part of , and a(v i ,e j )=0 means v i With e j Therefore, according to the definition of G, this study can construct two hypergraphs, namely and

[0099] S5. Generate high-order structural features of the network;

[0100] This step is mainly responsible for obtaining the feature representation of Chinese herbal medicines and indications through the hypergraph convolution algorithm, thereby mining potential associations.

[0101] The first step is to define the node initial vector. The node initial vector is the output of the low-dimensional feature acquisition module of the effective component and target, that is, the features of the effective component and target after dimensionality reduction.

[0102] The second step is hyperedge aggregation. The present invention uses a hypergraph convolutional neural network to learn the feature representation of hyperedges and mine potential associations. The hypergraph convolutional neural network can retain the local information of nodes through information transfer between neighbor nodes, and can also obtain the global information of nodes through information transfer and aggregation between non-neighbor nodes with common adjacent nodes. The convolution process for obtaining the features of Chinese herbal medicine and indications is the same. Here, Chinese herbal medicine is taken as an example. The first step is hyperedge aggregation. In this step, the present invention summarizes the information from V C Update ε with information C For each hyperedge in the hypergraph, the model collects information from its internal constituent nodes and then combines this information to generate a hyperedge embedding that captures high-order patterns and interactions between constituent nodes in the same hyperedge. For node-to-hyperedge conversion, the hyperedge aggregation is defined as:

[0103]

[0104] in, Yes A ch The edge degree matrix, E ch and M ch is the embedding matrix of active components and hyperedges at layer l. According to (6), the information transfer from an active component node to its hyperedge is The multiplication operation is implemented using The purpose is to simply rescale the hyperedge embeddings.

[0105] The third step is node aggregation. Using (6) we get M ch Finally, the present invention will be combined with E ch and M ch The embedding information of G is used to perform node aggregation to update the embedding of the active component nodes. This transformation process from hyperedges to nodes enables the present invention to capture the global context in G, which is defined as follows:

[0106]

[0107] in, Yes A ch The node degree matrix of is used to rescale the effective component embedding. Using (7), A ch With M ch Multiply them together to aggregate information from hyperedges to active component nodes.

[0108] Combining formulas (6) and (7), the complete hypergraph convolution can be defined as:

[0109]

[0110] However, when fusing multiple levels of aggregation during the learning process, the present invention may encounter the problem of feature information dilution. To this end, the present invention integrates jump connections similar to ResNet to avoid the information dilution problem caused by a large number of additional connections. The transformed hypergraph convolution is defined as follows:

[0111]

[0112] Among them, θ l is a trainable parameter, σ(·) is a nonlinear activation function, such as and With (9), the present invention can consider both the original features and the aggregated embeddings to generate the final representation of Chinese herbal medicine.

[0113] Replace A in (9) ch , and E ch Replace with A td , and E td , a similar hypergraph convolutional network can be constructed to learn the representation of indications through corresponding channels. Therefore, in the process of hypergraph convolution, through the two steps of hyperedge aggregation and node aggregation, the present invention can effectively capture and propagate information in the hypergraph structure. Considering the complex multi-component and multi-target relationship between Chinese herbal medicines and indications, this enables the present invention to learn meaningful and informative embeddings of Chinese herbal medicines and indications. These embeddings can be further used to predict the association between Chinese herbal medicines and indications.

[0114] Step 4: Model optimization. In order to make the learning process of the present invention more effective, cross-entropy (CE) loss is used during training, that is, It is a commonly used loss function that represents the error between the predicted result and the actual situation. is defined as follows:

[0115]

[0116] in, is the prediction score, l is the number of layers of the convolutional network, θ is the set of trainable parameters, and A hd It is the correlation matrix between Chinese herbal medicines and indications.

[0117] S6. Predict potential Chinese herbal medicine-indication associations;

[0118] The present invention finally uses a scoring function based on inner product to calculate the score of the association pair. In this process, the characteristic representations of the Chinese herbal medicine and indications obtained in the previous step are subjected to inner product operation to obtain the scores between different associations:

[0119]

[0120] In this formula, E H and E D They respectively represent the characteristics of Chinese herbal medicine and the characteristics of indications.

[0121] The score was compared with the set threshold, and a score higher than the threshold indicated that there was an association between the Chinese herbal medicine and the indication.

[0122] Example 2

[0123] The present application also provides a Chinese herbal medicine indication discovery system based on hypergraph convolution, which is implemented based on the above method and includes:

[0124] The data collection module is used to collect data on the relationship between Chinese herbal medicines, active ingredients, target sites and indications verified by biological experiments; obtain the structural data of active ingredients and the sequence information of target sites;

[0125] A data preprocessing module, used to calculate the fingerprint similarity of the active ingredients and generate an active ingredient similarity matrix, calculate the similarity of the target sequence and generate a target sequence similarity matrix based on the original data obtained by the data collection module;

[0126] The module for acquiring low-dimensional features of active ingredients and target points is used to reduce the dimension of the similarity matrix of active ingredients and target points by using an automatic encoder to acquire low-dimensional features of active ingredients and target points;

[0127] A hypergraph construction module is used to construct corresponding hypergraphs according to the collected Chinese herbal medicine-active ingredient association information and target point-indication association information;

[0128] The network high-order structural feature generation module is used to use the low-dimensional features of the active ingredients and target targets obtained by the low-dimensional feature acquisition module of the active ingredients and target targets as the initial features of the nodes in the hypergraph, and use two hypergraph convolutional networks to learn the high-order structural information in the Chinese herbal medicine-active ingredient and target target-indication hypergraphs, so as to obtain the feature representation of Chinese herbal medicine and indications in the network;

[0129] The potential Chinese herbal medicine-indication association prediction module is used to use the acquired feature representations of Chinese herbal medicines and indications as inputs to the scoring function to obtain scores for different Chinese herbal medicine-indication association pairs; and to determine whether there is an association between the Chinese herbal medicines and indications based on the scores.

[0130] The present application may also provide a computer device, comprising: at least one processor, a memory, at least one network interface and a user interface. The various components in the device are coupled together through a bus system. It is understood that the bus system is used to achieve connection and communication between these components. In addition to the data bus, the bus system also includes a power bus, a control bus and a status signal bus.

[0131] The user interface may include a display, a keyboard or a pointing device, such as a mouse, a trackball, a touch pad or a touch screen.

[0132] It is understood that the memory in the embodiments disclosed in the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memories.

[0133] In some embodiments, the memory stores the following elements, executable modules or data structures, or a subset thereof, or an extended set thereof: an operating system and applications.

[0134] The operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., which are used to implement various basic services and process hardware-based tasks. The application includes various application programs, such as a media player (Media Player), a browser (Browser), etc., which are used to implement various application services. The program for implementing the method of the embodiment of the present disclosure can be included in the application.

[0135] In the above embodiment, the processor may also call a program or instruction stored in the memory, specifically, a program or instruction stored in an application program, and is used to:

[0136] Execute the steps of the above method.

[0137] The above method can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor may be a general processor, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The above-disclosed methods, steps and logic block diagrams can be implemented or executed. The general processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the above-disclosed method can be directly embodied as a hardware decoding processor to execute, or the hardware and software modules in the decoding processor are combined to execute. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0138] It is understood that the embodiments described in the present application can be implemented by hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in the present application or a combination thereof.

[0139] For software implementation, the technology of the present application can be implemented by executing the functional modules (such as procedures, functions, etc.) of the present application. The software code can be stored in a memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0140] The present application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, each step in the above method embodiment can be implemented.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present application and are not intended to limit it. Although the present application is described in detail with reference to the embodiments, a person skilled in the art should understand that any modification or equivalent replacement of the technical solution of the present application does not depart from the spirit and scope of the technical solution of the present application and should be included in the scope of the claims of the present application.

Claims

1. A method for discovering indications of Chinese herbal medicines based on hypergraph convolution, comprising: S1: Collect data on the relationship between Chinese herbal medicines, active ingredients, target sites and indications verified by biological experiments; obtain the structural data of active ingredients and sequence information of target sites; S2: According to the original data obtained in step S1, the fingerprint similarity of the active ingredients is calculated and the active ingredient similarity matrix is ​​generated, and the similarity of the target sequence is calculated and the target sequence similarity matrix is ​​generated; S3: Use an autoencoder to reduce the dimension of the similarity matrix of the active ingredient and the target to obtain low-dimensional features of the active ingredient and the target; S4: construct corresponding hypergraphs based on the collected Chinese herbal medicine-active ingredient association information and target-indication association information; S5: using the low-dimensional features of the active ingredients and target points obtained in step S3 as the initial features of the nodes in the hypergraph, and using two hypergraph convolutional networks to learn the high-order structural information in the Chinese herbal medicine-active ingredients and target-indication hypergraphs, thereby obtaining the feature representation of Chinese herbal medicines and indications in the network; S6: Using the acquired characteristic representations of the Chinese herbal medicine and indications as inputs to the scoring function to obtain scores for different Chinese herbal medicine-indication association pairs; judging whether there is an association between the Chinese herbal medicine and the indication based on the scores.

2. The method for discovering Chinese herbal medicine indications based on hypergraph convolution according to claim 1, characterized in that: The step S2 comprises: Step S2-1: Convert the SMILES data of the active ingredient into a molecular fingerprint. When quantifying the similarity between two molecular fingerprints, the Dice coefficient is used to measure the overlap between the substructures represented by the fingerprints. Assume that c i and c j The Dice coefficient between two effective components is calculated as: Among them, c i and c j Represents two Chinese herbal medicine ingredients; M i and M j c i and c j The molecular fingerprint of To express the similarity between pairs of effective components, where n c is the number of components; Step S2-2: Obtain the sequence information of the target, calculate the similarity between the target sequences, and construct a similarity matrix, which is expressed as Where n t is the number of target points.

3. The method for discovering Chinese herbal medicine indications based on hypergraph convolution according to claim 3, characterized in that: The step S4 comprises: An active ingredient-Chinese herbal medicine hypergraph with active ingredients as nodes and Chinese herbal medicines as hyperedges and a target target-indication hypergraph with target targets as nodes and indications as hyperedges are constructed respectively.

4. The method for discovering Chinese herbal medicine indications based on hypergraph convolution according to claim 1, characterized in that: The step S5 comprises: Step S5-1: using low-dimensional features of active ingredients and target points as initial vectors of nodes of the hypergraph of active ingredients-Chinese herbal medicine and the hypergraph of target points-indications; Step S5-2: Learn the feature representation of hyperedges and mine potential associations through the hypergraph convolutional neural network, including: Step S5-2-1: Hyperedge aggregation: The hyperedge aggregation formula of the active ingredient-Chinese herbal medicine hypergraph is: in, is the indicator matrix A of the active ingredient-Chinese herbal medicine hypergraph ch The edge degree matrix of ch and M ch is the embedding matrix of active components and hyperedges at layer l; information from active component nodes to their hyperedges is transmitted through The multiplication operation is realized; the superscript T represents the matrix transpose; The hyperedge aggregation formula of the target-indication hypergraph is: in, is the indicator matrix A of the target-indication hypergraph td The edge degree matrix of td and M td is the target node and hyperedge embedding matrix of the lth layer; the information from the target node to its hyperedge is transmitted through The multiplication operation is realized; Step S5-2-2: Node aggregation: The node aggregation process of the active ingredient-Chinese herbal medicine hypergraph is expressed as: The conversion process from hyperedge to node is: in, A ch The node degree matrix of A is used to rescale the effective component embedding; ch With M ch Multiply,aggregate information from hyperedges to active component nodes; The complete hypergraph convolution is defined as follows: Among them, θ l is a trainable parameter; σ(·) is a nonlinear activation function; W C The encoding process of the automatic encoder in step S3: Among them, W enc is the weight matrix; b enc is the bias vector of the encoder; g(·) is the nonlinear activation function; For S C The i-th row vector of S C for the similarity between the active ingredients; The node aggregation process of the target-indication hypergraph is expressed as: The conversion process from hyperedge to node is: in, A td The node degree matrix of A is used to rescale the effective component embedding; td With M td Multiply, aggregate information from the hyperedge to the target node; The complete hypergraph convolution is defined as follows: in, W T The encoding process of the automatic encoder in step S3: in, For S T The i-th row vector of T is the similarity between the target sites; Step S5-2-3: Model training: using cross entropy as the loss function for model training.

5. The method for discovering Chinese herbal medicine indications based on hypergraph convolution according to claim 1, characterized in that: The step S6 comprises: Perform inner product operations on the characteristic representations of Chinese herbal medicines and indications in the network to obtain the scores between different associations Among them, E H and E D Represent the characteristics of Chinese herbal medicine and the characteristics of indications respectively; the superscript T indicates matrix transposition.

6. A Chinese herbal medicine indication discovery system based on hypergraph convolution, implemented based on any of the methods described in claims 1-5, characterized in that: The system comprises: The data collection module is used to collect data on the relationship between Chinese herbal medicines, active ingredients, target sites and indications verified by biological experiments; obtain the structural data of active ingredients and the sequence information of target sites; A data preprocessing module, used to calculate the fingerprint similarity of the active ingredients and generate an active ingredient similarity matrix, calculate the similarity of the target sequence and generate a target sequence similarity matrix based on the original data obtained by the data collection module; The module for acquiring low-dimensional features of active ingredients and target points is used to reduce the dimension of the similarity matrix of active ingredients and target points by using an automatic encoder to acquire low-dimensional features of active ingredients and target points; A hypergraph construction module is used to construct corresponding hypergraphs according to the collected Chinese herbal medicine-active ingredient association information and target point-indication association information; A network high-order structural feature generation module is used to use the low-dimensional features of the active ingredients and target targets obtained by the low-dimensional feature acquisition module of the active ingredients and target targets as the initial features of the nodes in the hypergraph, and use two hypergraph convolutional networks to learn the high-order structural information in the Chinese herbal medicine-active ingredient and target target-indication hypergraphs, so as to obtain the feature representation of the Chinese herbal medicine and indications in the network; and The potential Chinese herbal medicine-indication association prediction module is used to use the acquired feature representations of Chinese herbal medicines and indications as inputs to the scoring function to obtain scores for different Chinese herbal medicine-indication association pairs; and to determine whether there is an association between the Chinese herbal medicines and indications based on the scores.

Images