Structure data generation method, apparatus, device, and program

The method enhances molecular structure generation by using GCN-based feature extraction and iterative decoder training with wavelet transforms to address low diversity and efficiency issues, resulting in high-quality molecular structures for drug discovery.

JP2024541461AActive Publication Date: 2024-11-08TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
JP2024530434
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Priority Date
2022-02-17
Filing Date
2022-12-05
Publication Date
2024-11-08
Estimated Expiration
2042-12-05

AI Technical Summary

Technical Problem

Existing methods for generating molecular structures using graph convolutional networks (GCN) result in low diversity and efficiency due to low-pass filtering, leading to incomplete signal reconstruction and poor interpretability in decoding processes.

Method used

A method involving the acquisition of structural and node feature representations, generation of hidden layer features in multiple frequency bands using GCN, and iterative training of a decoder to reconstruct molecular structures, utilizing wavelet transforms for enhanced feature extraction and decoding.

Benefits of technology

Improves the efficiency and diversity of molecular structure generation by effectively reconstructing high-quality molecules with improved accuracy and adherence to chemical rules, enhancing drug discovery and development processes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024541461000001_ABST
    Figure 2024541461000001_ABST
Patent Text Reader

Abstract

The present invention relates to the field of artificial intelligence and provides a method, apparatus, device, medium and program product for generating structure data, which includes the steps of: obtaining a structural feature representation and a node feature representation of sample structure data (201); generating a hidden layer feature representation based on the structural feature representation and the node feature representation (202); inputting the hidden layer feature representation into a trained decoder to perform structural reconstruction to obtain predicted structure data (203); and training the trained decoder based on the predicted structure data to obtain a designated decoder (204), where the designated decoder is for obtaining reconstructed structure data by performing structural reconstruction on the inputted sample data.
Need to check novelty before this filing date? Find Prior Art

Description

Related Applications

[0001] This application claims priority to a Chinese patent application bearing application number 202210146218.2, filed on February 17, 2022, and entitled "Method, apparatus, device, medium and program product for generating structured data," the entire contents of which are incorporated herein by reference. [Technical field]

[0002] The present application relates to the field of artificial intelligence, and in particular to a method, an apparatus, a device, a medium and a program product for generating structured data. [Background technology]

[0003] With the development of artificial intelligence (AI), AI is being applied to more and more fields. Here, in the field of intelligent medicine, AI can promote drug discovery and assist experts in researching and developing new drugs.

[0004] In the related technology, chemical molecular structures are mapped to generate molecular graphs with corresponding graph structures, and then these molecular graphs are learned based on a message propagation process using a graph convolutional neural network (GCN). New feature representations are then generated using the GCN, and in the decision-making process, a new structure corresponding to the new feature representation is added to the existing graph so as to conform to the organic molecular chemistry rules, thereby obtaining a molecular graph corresponding to the new molecule.

[0005] However, in the process of generating the above-mentioned new molecular structures, the graph data representing the molecules is smoothed by the low-pass characteristics of the GCN, so it is not possible to reconstruct the complete original signal during decoding. Ultimately, the diversity and effectiveness of the generated molecules is lacking, resulting in low generation efficiency. Summary of the Invention [Problem to be solved by the invention]

[0006] The embodiments of the present application provide a method, an apparatus, a device, a medium, and a program product for generating structure data, which can improve the efficiency of generating a specified structure and the diversity of the generated structures. [Means for solving the problem]

[0007] According to one aspect of the present invention, a step of acquiring a structural characteristic representation and a node characteristic representation of sample structure data, the structural characteristic representation being for indicating a connection state between nodes constituting the sample structure data, and the node characteristic representation being for indicating a node type corresponding to a node constituting the sample structure data; generating hidden layer feature representations based on the structural feature representation and the node feature representation, the hidden layer feature representations being for indicating connectivity between nodes in the sample structural data in at least two frequency bands; obtaining predicted structure data by inputting the hidden layer feature representation to a trained decoder to perform structure reconstruction; The present invention provides a method for generating structure data, the method including: obtaining a designated decoder by training the training-awaiting decoder based on the predicted structure data, the designated decoder being for performing structure reconstruction on input sampling data to obtain reconstructed structure data, the sampling data being data obtained by sampling candidate data.

[0008] According to another aspect of the invention, an acquisition module for acquiring a structural characteristic representation and a node characteristic representation of sample structure data, the structural characteristic representation indicating a connection state between nodes constituting the sample structure data, and the node characteristic representation indicating a node type corresponding to the node constituting the sample structure data; an encoding module for generating hidden layer feature representations based on the structural feature representation and the node feature representation, the hidden layer feature representations for indicating connectivity between nodes in the sample structure data in at least two frequency bands; a decoding module for inputting the hidden layer feature representation to a trained decoder to perform structure reconstruction to obtain predicted structure data; and a training module for obtaining a designated decoder by training the training-awaiting decoder based on the predicted structure data, the designated decoder for performing structure reconstruction on input sampling data to obtain reconstructed structure data, the sampling data being data obtained by sampling candidate data.

[0009] According to another aspect of the present invention, there is provided a computing device comprising a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set which, when loaded and executed by the processor, realises a method for generating structure data as described in any one of the embodiments of the present application.

[0010] According to another aspect of the present invention, there is provided a computer-readable storage medium having stored thereon at least one program code which, when loaded and executed by a processor, realizes a method for generating structure data as described in any one of the embodiments of the present application.

[0011] According to another aspect of the present invention there is provided a computer program product or computer program comprising computer instructions stored on a computer readable storage medium for causing a processor of a computing device to read and execute the computer instructions from the computer readable storage medium, thereby causing the computing device to perform a method for generating structure data according to any of the above embodiments.

[0012] According to the technical solution of the present application, at least the following beneficial effects can be achieved: Effect of the Invention

[0013] After obtaining a hidden layer feature representation using a structural feature representation and a node feature representation corresponding to sample structural data, a designated decoder is obtained by iteratively training a decoder waiting to be trained using the hidden layer feature representation, thereby making it possible to generate required structural data using the sampling data input to the designated decoder. In other words, as necessary, a variety of reconstructed structural data can be quickly generated by the designated decoder obtained through training, thereby improving the efficiency and diversity of structural data generation. [Brief description of the drawings]

[0014] [Figure 1] FIG. 1 illustrates an implementation environment according to one exemplary embodiment of the present application. [Diagram 2] 1 is a flowchart of a method for generating structure data according to an exemplary embodiment of the present application; [Diagram 3] FIG. 2 illustrates the structure of a model according to one exemplary embodiment of the present application. [Figure 4] 1 is a flowchart of a method for generating hidden layer feature representations according to one exemplary embodiment of the present application. [Diagram 5] FIG. 2 illustrates the acquisition of hidden layer feature representation according to one exemplary embodiment of the present application. [Figure 6] FIG. 13 illustrates a diagram of obtaining a hidden layer feature representation according to another exemplary embodiment of the present application. [Figure 7] 1 is a flowchart of a method for training a decoder according to one exemplary embodiment of the present application; [Figure 8] 4 is a flowchart of a method for generating structure data according to another exemplary embodiment of the present application; [Figure 9] 4 is a flowchart of a method for generating structure data according to another exemplary embodiment of the present application; [Figure 10]FIG. 2 is a block diagram illustrating the structure of a structure data generating device according to an exemplary embodiment of the present application. [Figure 11] FIG. 2 is a block diagram showing the structure of a structure data generating device according to another exemplary embodiment of the present application; [Figure 12] FIG. 2 illustrates a structure of a server according to one exemplary embodiment of the present application. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] First, the terms used in the embodiments of the present application will be briefly explained.

[0016] Artificial Intelligence: Theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, sense the environment, acquire knowledge, and use the knowledge to obtain optimal results. In other words, artificial intelligence is a comprehensive technology of computer science that aims to understand the essence of intelligence and create new intelligent machines that can respond in a manner similar to human intelligence. In other words, artificial intelligence endows machines with the functions of sensing, reasoning, and decision-making by studying the design principles and implementation methods of various intelligent machines.

[0017] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware and software technologies. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed memory, big data processing technology, operation / interactive systems, and electro-mechanical integration. The software technology of artificial intelligence mainly includes several aspects such as computer vision technology, voice processing technology, natural language processing technology, and machine learning / deep learning, autonomous driving, and intelligent transportation.

[0018] Machine learning (ML) is an interdisciplinary field that involves many disciplines, such as probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It is a specialized study of how computers simulate or realize human learning behavior, thereby acquiring new knowledge and skills and reorganizing existing knowledge structures to continually improve their own performance. Machine learning is the core of artificial intelligence and the fundamental route to giving computers intelligence, and is applied across various fields of artificial intelligence. Machine learning and deep learning usually include techniques such as artificial neural networks, trust networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0019] Graph Convolutional Network: Convolutional neural network adopts local sensing region, shared weight value and downsampling in spatial domain, has stable characteristics against displacement, scaling and distortion, and can extract spatial features of images well. Graph structure does not have translation invariance of pictures, and traditional convolution methods cannot be applied to graph structure. Therefore, the main difficulty of graph convolution is that the number of neighbor nodes of each node in the graph does not match, and feature extraction cannot be performed with the same size of convolution kernel. GCN completes the integration of neighbor information through the message transmission mechanism of spatial or spectral domain to extract main features, but the most common GCN performs low-pass filtering on graph signals.

[0020] Wavelet Transform: It is a localized analysis of spatial frequency. It gradually performs multi-scale subdivision of the signal through stretching and translation operations, and finally reaches frequency band subdivision. It can automatically adapt to the requirements of time-frequency signal analysis, thereby focusing on any detail of the signal, and solving the difficult problems of Fourier transform.

[0021] Variational Auto-Encoder (VAE): A deep learning model for data generation. First, the input data is compressed and encoded, and then hidden variables are calculated and generated. Finally, the original data is restored by a decoder. During data generation, data close to the distribution of the original data can be generated simply by sampling from the hidden variables in a specific distribution. The VAE model is applied to molecule generation in order to generate effective molecules with properties consistent with reference molecules, and to discover higher quality drugs.

[0022] In an embodiment of the present application, machine learning / deep learning in artificial intelligence technology is applied to generate structured data that has or satisfies certain rules.

[0023] Next, an application scene of the structure data generating method according to the embodiment of the present application will be described as an example.

[0024] First, it can be applied to the generation of organic chemical molecules in intelligent medicine. In intelligent medicine, AI supports the discovery and research and development of new drugs, such as lead drug generation, drug optimization, etc., where the lead drug refers to a compound drug with certain activity and chemical structure obtained by a certain route and means, which is used for further structural modification and modification, and is the starting point of new drug research. Drug optimization refers to optimizing the structure of the chemical structure in the drug according to certain rules to improve the physicochemical properties of the drug.

[0025] In the related technology, the chemical molecules of a compound are mapped to a molecular graph of a graph structure, and then feature extraction is performed by a graph convolutional neural network (GCN), and then the structure is restored by the GCN, that is, the intermediate feature Z=GCN(X,A) is obtained by feature extraction, where X is the node feature of the molecular graph, and A is the edge feature of the molecular graph, and then TIFF2024541461000002.tif6170 is generated, TIFF2024541461000003.tif6170 is the characteristic after the specified conversion is performed on A. However, in the process of realizing this method, the interpretability of the decoding method used in the process of generating new molecules is poor, and there is no decoupling principle that is dual to the encoding part, so secondary smoothing is performed on the graph signal, and the diversity of the generated molecular graph is low, and the generation efficiency is also low.

[0026] For example, a designated decoder can be obtained by learning the chemical molecular structure of an existing drug using the structural data generation method according to the embodiment of the present application, and since the decoding process and the encoding process are dual, the designated decoder can efficiently and effectively generate high-quality clinical candidate molecules in the research and development process of new drugs to support the research and development of new drugs. Alternatively, the method can be applied to the screening process of similar drug molecules with relatively strong potential activity on known targets, where a similar drug molecule is a compound corresponding to a molecule that has a certain similarity to a known drug and has the potential to become a drug, and the structural data generation method according to the embodiment of the present application generates an ideal candidate molecule with a high success rate for targets that are difficult to become drugs.

[0027] Second, it can be applied to the scene of mining and construction of knowledge graphs. Here, the knowledge graph is composed of several entities and their attributes connected to each other, and is a modern theory that combines theories and methods of applied mathematics, graphics, information visualization technology, information science, etc. with methods such as metrology citation analysis and co-occurrence analysis to visualize the core structure, advanced fields, and the entire knowledge architecture of the disciplines to achieve the purpose of multi-disciplinary integration by using visualized graphs. Specifically, for example, the medical knowledge graph in the intelligent medical scene is illustratively obtained by training a knowledge graph corresponding to a previously constructed medical condition as training data to obtain a designated decoder, and efficiently generate a number of knowledge graphs with a certain validity by the designated decoder.

[0028] Third, it can be applied to the scene of intelligent automatic travel planning. For example, according to the method for generating structure data according to the embodiment of the present application, a designated decoder is obtained by training according to a travel itinerary planning graph for training, and a user can generate various travel planning courses according to the designated decoder, that is, a variety of travel itinerary plans can be provided to the user according to a designated condition or a random condition, and intelligent travel can be enriched, and the above-mentioned designated condition can be a designated travel city, a designated tourist attraction type, etc.

[0029] The above exemplary scenes are merely examples of application scenes of the method for generating structured data according to the embodiments of the present application. The method may be applied to scenes in which information such as a recommendation system based on social relationships between users, text semantic analysis, and road condition prediction can be processed into graph-structured data, and the specific application scenes are not limited here.

[0030] The implementation environment of the embodiment of the present application will be described in combination with the above description of the noun interpretation and application scene. As shown in FIG. 1, the computer system of the implementation environment includes a terminal device 110, a server 120, and a communication network 130.

[0031] The terminal device 110 includes various types of devices, such as mobile phones, tablet computers, desktop computers, mobile notebook computers, smart home appliances, in-vehicle terminals, airplanes, etc. Illustratively, a user instructs the server 120 to train a training-waiting decoder through the terminal device 110.

[0032] The server 120 is for providing a training function for the decoder waiting to be trained, that is, the server 120 can call a corresponding computing module to train a specified decoder waiting to be trained according to a request of the terminal device 110. Preferably, a model architecture corresponding to a decoder waiting to be trained may be pre-stored in the server 120, or may be uploaded by the terminal device 110 through a model data file. A training dataset used for training the decoder waiting to be trained may be pre-stored in the server 120, or may be uploaded by the terminal device 110 through a training data file. In one example, a user uploads a dataset corresponding to sample structure data to the server 120 through the terminal device 110, and sends a training request for the decoder waiting to be trained, which includes a model identifier (ID) of the decoder waiting to be trained, and the server 120 reads the model architecture of the decoder waiting to be trained corresponding to the model ID from a database according to the model ID in the training request, and trains the decoder waiting to be trained with the received dataset.

[0033] Here, in the training process, the server 120 obtains a hidden layer feature representation according to the structural feature representation and the node feature representation of the sample structural data, and trains the decoder waiting to be trained according to the hidden layer feature representation to obtain a designated decoder. After the server 120 performs training and obtains a designated decoder, the server 120 can send the designated decoder to the terminal device 110, or set the designated decoder in an application module so that the terminal device 110 can call it by a data generation request.

[0034] In some embodiments, if the computing power of terminal device 110 is sufficient for the training process of the training-awaiting decoder, the entire training process of the designated decoder may be implemented independently by terminal device 110.

[0035] In addition, the server 120 may be an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Networks (CDNs), and big data and artificial intelligence platforms.

[0036] Here, cloud technology refers to a trusted technology that integrates resources such as hardware, software, and networks within a wide area network or a local area network to realize data calculation, storage, processing, and sharing.

[0037] In some embodiments, the server 120 may be implemented as a node in a blockchain system.

[0038] For example, the terminal device 110 and the server 120 are connected via a communication network 130, where the communication network 130 may be a wired network or a wireless network, and is not limited thereto.

[0039] FIG. 2 illustrates a method for generating structure data according to one embodiment of the present application, in which the method is executed by a computing device, which may be implemented as a terminal device or a server in FIG. 1. In one example, the method is applied to the server shown in FIG. 1 and includes the following steps:

[0040] In step 201, a structural feature representation and a node feature representation of the sample structure data are obtained.

[0041] Here, the structural characteristic representation is intended to indicate the connection status between the nodes constituting the sample structure data, and the node characteristic representation is intended to indicate the node type corresponding to the nodes constituting the sample structure data.

[0042] Exemplarily, the sample structure data is training data for training a training decoder, and the sample structure data is data whose data structure is a graph structure, that is, the sample structure data is data consisting of at least two nodes and at least one edge.Preferably, the graph structure may be any of the graph structures of an undirected graph, a directed graph, an undirected complete graph, a directed complete graph, etc., and the specific graph structure can be determined based on the data information corresponding to the sample structure data.For example, when it is necessary to represent a chemical molecule by a graph structure, that is, when the sample structure data corresponds to a chemical molecule, the atoms in the molecule become nodes in the graph, and the chemical bonds between the atoms become edges in the graph, and the edges do not need to indicate a direction, so that an undirected graph can be used as the data structure corresponding to the sample structure data.

[0043] Here, the structural feature representation is for indicating the connection status between nodes in a graph corresponding to the sample structural data, and the connection status is related to the structure generation task corresponding to the sample structural data. For example, if the structure generation task is the generation of chemical molecules, the connection relationship between the nodes is the chemical bond between each atom in the chemical molecule. If the structure generation task is a recommendation system based on a social network, the connection status between the nodes is the mutual relationship between users in the social network (e.g., stranger relationship, friend relationship, blacklist relationship, etc.). If the structure generation task is the generation of a travel course, the connection status between the nodes is the route status between tourist spots.

[0044] The node feature representation is for indicating a node type of each node in the graph corresponding to the sample structure data, and the node type is related to the structure generation task corresponding to the sample structure data. For example, if the structure generation task is generation of a chemical molecule, the node type is an atom type in the chemical molecule. If the structure generation task is a recommendation system based on a social network, the node type is a user account in a social network. If the structure generation task is generation of a travel course, the node type is a tourist attraction.

[0045] Exemplarily, the structural feature representation and the node feature representation of the sample structure data are obtained by converting the sample structure data according to a preset feature conversion manner. Preferably, the structural feature representation may be a matrix-type feature representation or a vector-type feature representation, and the node feature representation may be a matrix-type feature representation or a vector-type feature representation, which is not limited herein.

[0046] Preferably, the preset transformation method may be a network transformation method, i.e., the structural feature representation and the node feature representation may be obtained by performing feature extraction by a feature extraction network, and the structural feature representation and the node feature representation are obtained by inputting sample structural data into a pre-trained feature extraction network and outputting the structural feature representation and the node feature representation. The feature extraction network may be a network capable of completing feature extraction, such as a Convolutional Neural Network (CNN), a Visual Geometry Group Network (VGGNet), an Alex Network (AlexNet), etc., but is not limited thereto.

[0047] Preferably, the above-mentioned preset transformation method may be a matrix transformation method, that is, the above-mentioned structure characteristic representation and node characteristic representation may be obtained by matrix transformation of the graph structure data corresponding to the sample structure data. Illustratively, the logical structure of the sample structure data of the graph structure is divided into two parts: a node set consisting of at least two nodes and an edge set consisting of edges between the nodes. The above-mentioned structure characteristic representation is an adjacency matrix, which is two-dimensional data generated by the edge set and stores edges, and the adjacency matrix is ​​for recording the connection relationship between the at least two nodes, and the above-mentioned node characteristic representation is a one-dimensional matrix (array) generated by the node set and for storing node data in a graph.

[0048] In one example, take the sample structure data as an example for indicating a chemical molecule, the recorded in the above-mentioned adjacency matrix is ​​the chemical bond type between atoms in the chemical molecule, and the node feature representation is a one-dimensional feature matrix generated based on the constituent atoms of the chemical molecule for recording the atom types in the chemical molecule. Illustratively, a sample chemical molecule consisting of at least two atoms is obtained, and the sample chemical molecule is a known molecule that satisfies the atomic bond criteria. The sample chemical molecule is converted into a sample molecular graph whose data structure is a graph structure, and the nodes of the sample molecular graph are for representing at least two atoms in the sample chemical molecule, such as carbon atom, hydrogen atom, oxygen atom, etc., and the edges in the sample molecular graph are for indicating chemical bonds between atoms in the sample chemical molecule, and the above-mentioned chemical bonds include types such as edgeless, single bond, double bond, triple bond, etc., in which the atom type and chemical bond type can be mapped to different characters or strings according to a specified mapping relationship in a computer, for example, edgeless can be corresponding to "0", single bond can be corresponding to "1", double bond can be corresponding to "2", and triple bond can be corresponding to "3", respectively, and the above-mentioned mapping relationship can be recorded in a pre-set mapping table. The adjacency matrix corresponding to the sample molecular graph is determined as a structural feature representation, and the node matrix corresponding to the sample molecular graph is determined as a node feature representation. Since the molecular graph of the graph structure can simply and clearly represent the connection relationship between atoms in the chemical molecule, the efficiency of obtaining the sample molecular graph is improved, and the features of the atom type and the features of the chemical bond between atoms can be reserved in the process of feature extraction.

[0049] In the embodiment of the present application, the above-mentioned training-awaiting decoder is a part in a training model. Illustratively, the structural feature representation and the node feature representation corresponding to the sample structural data are input into the training model to output the predicted structural data, and the whole training model is trained according to the deviation between the predicted structural data and the sample structural data, that is, the training of the training-awaiting decoder is completed in the whole training process for the training model.

[0050] In step 202, a hidden layer representation is generated based on the structural and node representations.

[0051] The hidden layer feature representation is intended to indicate connectivity between nodes in the sample structure data in at least two frequency bands.

[0052] Exemplarily, the training model further includes a training-awaiting encoder, which is for generating hidden layer feature representation according to the structural feature representation and the node feature representation.Preferably, the encoder structure corresponding to the training-awaiting encoder can be an autoencoder, a variational autoencoder (VAE), a low-pass filter, a band-pass filter, etc., and the specific filter used can be a convolutional neural network, a wavelet filter, a Butterworth filter, a Bessel filter, etc., which are not limited here.

[0053] In the embodiment of the present application, when the sample structure data is graph structure data, the trained encoder is a GCN, that is, the structure feature representation and the node feature representation are input to the GCN, and the hidden layer feature representation is output.

[0054] Illustratively, since the hidden layer feature representation is for indicating the connectivity between nodes in the sample structure data in at least two frequency bands, the GCN in the embodiment of the present application includes a first filtering layer and at least two second filtering layers.

[0055] Preferably, the first filtering layer may be for completing low-pass filtering, may be for completing high-pass filtering, or may be for completing band-pass filtering, and specifically needs to be determined according to the needs of the actual application scene. In one example, when applied to the generation scene of chemical molecules, the first filtering layer is a low-pass GCN layer. Here, the first filtering layer only indicates the filtering layer with matching function, and the first filtering layer may be composed of a single filtering layer, or may be composed of multiple filtering layers, for example, the first filtering layer includes two low-pass GCN layers, that is, composed of two layers of neurons, and is not limited here.

[0056] The second filtering layer is a bandpass filtering layer corresponding to at least two frequency bands, and exemplarily, the data output from the first filtering layer is input to the at least two second filtering layers, and the filtering results corresponding to each frequency band are output. Exemplarily, the number of second filtering layers corresponds to the division number of the frequency band. Here, the second filtering layer merely indicates a filtering layer having a matching function, and the at least two second filtering layers are at least two second filtering layers connected in parallel, and one second filtering layer may be composed of a single filtering layer or may be composed of multiple filtering layers, which is not limited here.

[0057] Illustratively, intermediate feature data is obtained by respectively encoding at least two frequency bands based on the structure feature representation and the node feature representation, the intermediate feature data is for indicating the coupling situation between nodes of the sample structure data in the corresponding frequency band, and the intermediate feature data corresponding to the at least two frequency bands respectively are clustered based on the designated data distribution to obtain a hidden layer feature representation, that is, the intermediate feature data in at least two frequency bands is obtained by the first filtering layer and the second filtering layer, and the hidden layer feature representation is obtained by clustering the intermediate feature data.

[0058] Preferably, the clustering method of the intermediate feature data may be splicing based on a frequency band order between at least two frequency bands, and in one example, the frequency band order may be an order in which the at least two frequency bands are arranged from low frequency to high frequency. Alternatively, the clustering method of the intermediate feature data may be fitting the intermediate feature data corresponding to at least two frequency bands based on a specified data distribution, for example, fitting based on a normal distribution (Gaussian distribution), fitting by Chebyshev polynomials, fitting based on the least squares method, etc. In some embodiments, when the number of nodes in the sample structure data is small, that is, when the computation power can meet the computation demands of the structure matrix and the feature matrix, decomposition of the feature value may be used instead of polynomial fitting, that is, feature decomposition may be performed on the Laplacian matrix.

[0059] In step 203, the hidden layer feature representation is input to a trained decoder to perform structure reconstruction, thereby obtaining predicted structure data.

[0060] Exemplarily, the hidden layer feature representation is reconstructed by the decoder to obtain the predicted structure data. In some embodiments, the output of the train-awaiting decoder is a decoded structure feature representation and a decoded node feature representation, i.e., the train-awaiting decoder obtains the decoded structure feature representation and the decoded node feature representation by sampling the hidden layer feature representation, and generates the predicted structure data based on the decoded structure feature representation and the decoded node feature representation, where the decoded structure feature representation is for indicating the relationship between nodes in the predicted structure data, and the decoded node feature representation is for indicating the nodes in the predicted structure data.

[0061] In one example, Figure 3 is a diagram illustrating a model structure according to one exemplary embodiment of the present application. As shown in Figure 3, the structural feature representation and the node feature representation of the sample structural data are input to a first filtering layer 310, and then the output result of the first filtering layer 310 is input to a second filtering layer 320, and the output of the second filtering layer 320 is clustered to obtain a hidden layer feature representation, and the hidden layer feature representation is decoded by a decoder 330 to obtain a decoded result, which is the decoded structural feature representation and the decoded node feature representation.

[0062] In step 204, a designated decoder is obtained by training a training-waiting decoder based on the prediction structure data.

[0063] Here, the designated decoder is for reconstructing the structure of the input sampled data to obtain reconstructed structure data, and the sampled data is data obtained by sampling candidate data. In some embodiments, the candidate data is data satisfying a designated data distribution.

[0064] In the embodiments of the present application, the trained decoder is trained according to the structural difference situation between the predicted structural data output from the trained decoder and the input sample structural data until the trained decoder converges. In some embodiments, the trained decoder is a part of the training model, so its training process depends on the overall training process of the training model, that is, the trained model is trained according to the structural difference situation between the predicted structural data and the input sample structural data to obtain a converged prediction model, and the decoder part in the prediction model is decomposed into the designated decoder, which is used to generate the reconstruction structural data.

[0065] For example, a training loss value is obtained based on the structural difference situation between the sample structural data and the predicted structural data, and in response to the training loss value reaching a designated loss threshold, i.e., in response to the determination that the model has been trained until convergence, the training of the decoder waiting to be trained is determined to be completed, and a designated decoder is obtained. Or, in response to the failure of matching the training loss value with the designated loss threshold, iterative training is performed on the model parameters of the decoder waiting to be trained, i.e., iterative training is performed by adjusting the model parameters of the training model. Here, the designated loss threshold may be preset by the system or customized according to the needs of the user, for example, the higher the model accuracy requirement required, the smaller the corresponding designated loss threshold.

[0066] The training loss value is calculated by a specified loss function, which may be a loss function used in regression, reconstruction and classification, and preferably, the specified loss function may be a loss function such as a mean absolute error loss function, a negative log-likelihood loss function, an exponential loss function, a cross-entropy loss function and variants thereof, but is not limited thereto.

[0067] As described above, the method for generating structure data in the embodiments of the present application obtains a hidden layer feature representation using a structural feature representation and a node feature representation corresponding to sample structure data, and then performs iterative training on a decoder waiting to be trained using the hidden layer feature representation to obtain a designated decoder, thereby making it possible to generate required structure data using the sampling data input to the designated decoder. In other words, as necessary, various reconstructed structure data can be quickly generated by the designated decoder obtained through training, thereby improving the efficiency and diversity of generation of structure data.

[0068] 4 shows a method for generating hidden layer feature representation according to an embodiment of the present application, in which the process of obtaining hidden layer feature representation by an encoder is described. The method includes the following steps:

[0069] In step 401, a structural feature representation and a node feature representation of the sample structure data are obtained.

[0070] In the present embodiment, the overall training model includes an encoder portion used for estimation and a decoder portion used for generation, and the structural feature representation and the node feature representation of the sample structure data are the inputs of the overall training model.

[0071] Here, the structural feature representation is for indicating a connection state between nodes constituting the sample structure data, and the node feature representation is for indicating a node type corresponding to the node constituting the sample structure data. The sample structure data is training data for training a training model, and the sample structure data is data whose data structure is a graph structure.

[0072] In step 402, intermediate feature data is obtained by encoding in at least two frequency bands based on the structural feature representation and the node feature representation, respectively.

[0073] In an embodiment of the present application, intermediate feature data is obtained by encoding the structural feature representation and the node feature representation by a training-awaiting encoder in a training model, where the training-awaiting encoder structure includes a first filtering layer and at least two second filtering layers.

[0074] In some embodiments, the encoding process is completed by wavelet transform, that is, the structure of the above-mentioned training-waiting encoder is a GCN, where the first filtering layer is a low-pass GCN layer, and the second filtering layer is a band-pass wavelet layer, where each of the at least two band-pass wavelet layers corresponds to a different wavelet basis function, that is, the signal is filtered by the wavelet basis function, that is, referring to the multi-scale principle in wavelet transform, it represents the basis of convolution band-pass filtering based on the Taylor expansion formula of different wavelet basis functions, and the wavelet transform process is a convolution process between the input features and the wavelet basis function.

[0075] Illustratively, a scale criterion is obtained based on a structure generation task to be completed by the training model, and at least two corresponding basis functions are calculated based on the scale criterion, and the at least two basis functions form a basis function set, and each basis function in the basis function set corresponds to one bandpass wavelet layer, i.e., corresponds to one frequency band.

[0076] In step 403, a hidden layer feature representation is obtained by clustering the intermediate feature data corresponding to each of the at least two frequency bands based on a specified data distribution.

[0077] In some embodiments, the outputs of at least two band-pass wavelet layers are directly clustered to obtain hidden layer feature representations, that is, in this case, the intermediate feature data is the encoding result obtained by the trained encoder, and for example, the structure feature representation and the node feature representation are input to the trained encoder, and intermediate feature data corresponding to at least two frequency bands are output, and the intermediate feature data is clustered to obtain hidden layer feature representations. That is, in the process of generating hidden layer feature representations for the structure feature representation and the node feature representation, the intermediate feature data is generated by the trained encoder, and the trained encoder and the trained decoder are jointly trained, so that the training result is back-propagated in the model training process to support the training of the trained decoder.

[0078] Here, the clustering method of the intermediate feature data splices based on a frequency band order between at least two frequency bands, and in one example, the frequency band order may be an order in which the at least two frequency bands are arranged from low frequency to high frequency.

[0079] Illustratively, FIG. 5 illustrates a diagram of obtaining hidden layer feature representations according to one exemplary embodiment of the present application, where the structure feature representation and the node feature representation 501 are input to the GCN0 510 to obtain low-pass filtered data 502, and the low-pass filtered data 502 is then input to the GCN0 510 to obtain the hidden layer feature representation. wavelet 520(GCN wavelet1 , G.C.N. wavelet2 , … , G.C.N. waveletn Enter the GCN wavelet Each of the nodes 520 outputs one intermediate feature data 503, and the intermediate feature data 503 is clustered to obtain a hidden layer feature representation Z504.

[0080] In some other embodiments, the intermediate feature data may be obtained by performing intermediate calculations on the features encoded by the trained encoder, in order to provide a diversity of the reconstruction structure data generated by the trained designated decoder.

[0081] In one example, the data obtained by the above intermediate calculation is the mean value and variance. Illustratively, based on the structural feature representation and the node feature representation, the node feature vectors of the nodes of the sample structural data in the feature space corresponding to at least two frequency bands are obtained, and the mean value data and variance data between the node feature vectors corresponding to at least two frequency bands are obtained, and the mean value data and variance data are determined as intermediate feature data. That is, after filtering by the trained encoder, the mean value and variance are calculated for the node feature vectors corresponding to each frequency band to obtain the mean value data and variance data, and the mean value data and variance data are intermediate feature data for generating the hidden layer feature representation. That is, by realizing the variation estimation process of the variational autoencoding based on the mean value data and variance data between the node feature vectors corresponding to different frequency bands, the interpretability of the designated decoder obtained by training downstream is improved, and the generation space of the structural data is expanded.

[0082] Specifically, the trained encoder is an encoder based on the probability model shown in Equation 1, where Z represents the hidden layer feature representation of the node, X represents a feature matrix corresponding to the node feature representation, the dimension of X is N*1, N is the quantity of nodes in the sample structure data, N is a positive integer, and A is an adjacency matrix for storing edges between nodes corresponding to the structure feature representation.

[0083]

number

[0084] Here, q(z i |X,A) is determined by Equation 2, where μ represents the mean value of the node vector representation, σ represents the variance of the node vector representation, and diag() is to generate a diagonal matrix. Equation 2 is i represents fitting the features corresponding to a Gaussian distribution.

[0085]

number

[0086] As can be seen from Equation 1 and Equation 2, when the structure feature representation and the node feature representation are known conditions, the training waiting encoder determines the connection probability that the i-th prediction node in the prediction node set establishes a connection relationship with each prediction node in the prediction node set, and determines the connection probability distribution corresponding to the i-th prediction node according to the connection probability between the i-th prediction node and each prediction node, and determines intermediate feature data (distribution corresponding to hidden layer feature representation) corresponding to at least two frequency bands respectively according to the fusion result of the connection probability distributions of all prediction nodes in the prediction node set, where the prediction node is for constructing the prediction structure data to be finally output. Exemplarily, each prediction node in the prediction node set corresponds to one connection probability distribution, and the distribution obtained by continuously multiplying the connection probability distributions corresponding to each prediction node can be determined as the distribution corresponding to the hidden layer feature representation Z, and in the embodiment of the present application, the distribution of the frequency band hidden layer feature representation corresponds to at least two frequency bands, and the distribution corresponding to the hidden layer feature representation can be obtained by fusion of the distributions of the frequency band hidden layer feature representations of the above different frequency bands.

[0087] For example, the structural data is a chemical molecule. A set of different types of chemical atoms is a prediction node set. The i-th type of chemical atom is associated with a probability value of establishing a connection relationship with all types of chemical atoms in the prediction node set, including the i-th type of chemical atom, and the connection relationship between the i-th type of chemical atom and all types of chemical atoms in the set is comprehensively observed to form the above-mentioned connection probability distribution, which is a distribution of the above-mentioned probability values ​​corresponding to the i-th type of chemical atom. Each chemical atom corresponds to a connection probability distribution, and the connection probability distribution is fitted to a Gaussian distribution during the training process of the training encoder. When the connection probability distribution of each node is fitted to a Gaussian distribution, the distribution of the hidden layer feature representation Z obtained by successively multiplying the connection probability distribution is also a Gaussian distribution.

[0088] If GCN is adopted in the above probability model, the mean value data and variance data can be obtained by encoding. Take single-layer GCN as an example, the calculation expression of single-layer GCN is as shown in Equation 3, where A is the structural feature representation input to GCN, X is the node feature representation input to GCN, W0 is the model parameter of GCN model, and D is the degree matrix corresponding to the sample structure data of graph structure. Preferably, the mean value data and variance data can be obtained by single-layer GCN or multi-layer GCN, and this is not limited here.

[0089]

number

[0090] After obtaining the mean data and the variance data, the mean data and the variance data can be fitted based on a specified data distribution to obtain a hidden layer feature representation. In one example, the specified data distribution is a Gaussian distribution.

[0091] 6 is a diagram illustrating the acquisition of hidden layer feature representations according to another exemplary embodiment of the present application. The structure feature representation and the node feature representation 601 are input to the GCN0 610 to obtain low-pass filtered data 602, and the low-pass filtered data 602 is input to the GCN0 610. wavelet 620(GCN wavelet1 , G.C.N. wavelet2 , … , G.C.N. waveletn Enter the GCN wavelet A corresponding node feature vector 603 is output from each of 620, and mean value data and variance data 604 corresponding to the node feature vector 603 are obtained by intermediate calculation. The mean value data and variance data 604 are used as intermediate feature data, and fitting is performed based on the intermediate feature data to obtain a hidden layer feature representation Z605 of a Gaussian distribution.

[0092] In the present embodiment, for GCN0 shown in FIG. 6, the corresponding calculation is as shown in Equation 3, i.e. TIFF2024541461000007.tif6170. The node feature representation X and the structure feature representation A are input to GCN0, and are transformed by the activation function to obtain X0. wavelet The inputs are X0 and A above, and GCN wavelet GCN wavelet1 , G.C.N. wavelet2 For example, if you include GCN wavelet1 (X0,A)=A wavelet1 X0W1, GCN wavelet2 (X0,A)=A wavelet2 X0W2, where W1 is the GCN wavelet1 are the model parameters of GCN wavelet2 are the model parameters of A wavelet1 GCN wavelet1 is the wavelet transform result of A with the wavelet basis function corresponding to A wavelet2 GCN wavelet2Here, the activation function may be a sigmoid activation function (S-shaped growth curve), a tanh nonlinear function, a ReLU activation function, or a variant thereof, and is not limited thereto.

[0093] In one example, after determining the mean data and variance data, the hidden layer feature representation is calculated using the reparameterization trick to obtain the hidden layer feature representation, i.e., as shown in Equation 4, where μ is the mean data, σ is the variance data, and ε is a normal Gaussian distribution, i.e., p(ε)=N(0,I).

[0094]

number

[0095] As described above, the method for generating hidden layer feature representations according to the embodiments of the present application uses GCN to filter the structure feature representations and node feature representations corresponding to sample structure data to obtain intermediate feature data corresponding to multiple frequency bands, and the intermediate feature data is clustered to obtain hidden layer feature representations, that is, the process of graph compression encoding is realized by wavelet transform, the feature is subjected to multi-scale subdivision to achieve multi-frequency banding of the feature, and the feature is focused on the feature details corresponding to multiple frequency bands respectively, thereby ensuring the diversity of the generated reconstructed structure data. At the same time, the encoding prediction process is realized using graph structure data to ensure the accuracy requirements of data for reconstructing the structure.

[0096] FIG. 7 illustrates a method for training a decoder according to one embodiment of the present application, in which the embodiment of the present application describes a decoder portion in a training network, where steps 701 to 703 (including steps 7031 and 7032) are implemented after step 403, and the method includes the following steps:

[0097] In step 701, a trained decoder performs structural reconstruction on the hidden layer feature representation to obtain a decoded structural feature representation and a decoded node feature representation.

[0098] In the embodiment of the present application, the hidden layer feature representation Z is the input of the trained decoder, and the trained decoder reconstructs the decoded structure feature representation and the decoded node feature representation according to the probability that edges exist between the nodes in the decoding process, where the probability model corresponding to the decoder is as shown in Equation 5, where N is the quantity of nodes, and z i and z j are nodes in the hidden layer feature representation.

[0099]

number

[0100] Here, p(A i,j |z i ,z j ) is given by Equation 6, where sigmoid() represents the activation function (S-shaped growth curve), TIFF2024541461000010.tif6170 is z i The sigmoid activation function used above is merely an example, and in actual applications, other activation functions may be used, which is not limited here.

[0101]

number

[0102] As can be seen from the above formulas 5 and 6, the obtained TIFF2024541461000012.tif6170 is as shown in Equation 7, where Z is the hidden layer feature representation and σ() is the activation function similar to sigmoid() above.

[0103]

number

[0104] In the embodiment of the present application, the trained decoder reconstructs the predicted structure data according to the inverse transform theory of the wavelet base to obtain the decoded structure feature representation and the decoded node feature representation. Here, the process of completing the reconstruction of the structure by the wavelet inverse transform can also be extended to some other high frequency basis functions, such as high pass filtering function, any wavelet base that does not have a generating function, etc.

[0105] In the decoding process, it is necessary to discretize the scale in the wavelet transform process. In one example, the inverse transform expression of the kernel function g is defined as g-1, and the inverse function at a=1 is TIFF2024541461000014.tif6170, inverse function at a=2 TIFF2024541461000015.tif6170, inverse function at a=3 TIFF2024541461000016.tif6170 are obtained correspondingly, where a is the scale in the wavelet basis function. Then, the inverse representation coefficient is obtained by performing a third-order Taylor expansion on the three inverse functions. Here, the above scale division (i.e., a=1, 2, 3) is merely exemplary, and in actual applications, scale division may be performed in different ways, which is not limited here. The hidden layer feature representation is convoluted with the inverse representation coefficient to obtain the decoding structure feature representation and the decoding node feature representation.

[0106] In step 702, predicted structure data is generated based on the decoded structure feature representation and the decoded node feature representation.

[0107] For example, the connection relationship between nodes in the predicted structure data can be known by obtaining the decoded structure feature representation, and the nodes in the predicted structure data can be known by obtaining the decoded node feature representation, and the predicted structure data of the corresponding graph structure can be obtained according to the nodes and the connection relationships between the nodes. The predicted structure data is for participating in the training process of the training model as the output result of training.

[0108] In one example, the decoded structure feature representation is realized as an adjacency matrix obtained by decoding, and the decoded node feature representation is realized as a node matrix obtained by decoding, where the adjacency matrix is ​​a matrix for representing the adjacency relationship between nodes, and the node matrix is ​​for indicating the node type corresponding to each node in the predicted structure data. The matrix element in the i-th row and j-th column in the adjacency matrix represents the connection status between the i-th node and the j-th node, for example, when the matrix element in the i-th row and j-th column is 0, it represents that there is no edge between the i-th node and the j-th node. Based on the adjacency matrix, the connection status between each node in the node matrix can be determined, and predicted structure data can be further generated.

[0109] In step 7031, in response to a failure to match the training loss values ​​between the sample structure data and the predicted structure data with a specified loss threshold, model parameters of the training model are iteratively trained.

[0110] For example, obtain the training loss value according to the structural difference situation between the sample structure data and the predicted structure data. In the embodiment of the present application, in the training process, determine the training loss value according to both the distance metric between the generated predicted structure data and the originally input sample structure data and the divergence between the node distribution and the Gaussian distribution, that is, obtain the distance metric data of the sample structure data and the predicted structure data in the feature space, obtain the divergence data between the node distribution corresponding to the predicted structure data and the designated data distribution, the node distribution is for indicating the distribution situation of the node feature vector of the predicted structure data in the feature space, and the divergence data is for indicating the degree of difference between the node distribution and the designated data distribution, and obtain the training loss value according to the distance metric data and the divergence data. That is, the feature similarity between the sample structure data and the predicted structure data in the feature space is indicated based on the distance metric data between the two, and the training loss value used for model training is jointly determined based on the divergence between the node distribution of the predicted structure data and the specified data distribution, thereby training the model from two angles, that is, feature difference and node distribution, and improving the prediction accuracy of the structure data by the model during the training process, and fitting the node distribution to the specified data distribution as much as possible, and sampling the specified data distribution downstream to generate corresponding reconstructed structure data, thereby improving the diversity of the generated structure data.

[0111] Here, the training loss value is calculated by a specified loss function, which may preferably be a loss function such as mean absolute error loss function, negative log-likelihood loss, exponential loss, cross entropy loss function and variants thereof.

[0112] Preferably, the distance metric data may be Euclidean distance, Hamming distance, cosine similarity, Manhattan distance, Chebyshev distance, etc., of the sample structure data and the predicted structure data in the feature space, but is not limited thereto.

[0113] In one example, the specified loss function for determining the training loss value is as shown in Equation 8, where: TIFF2024541461000017.tif6170 is the cross entropy loss function between the structural features and the node features, p(Z) is as shown in Equation 9, and KL() is the relative entropy (Kullback-Leibler) divergence function.

[0114]

number

number

[0115] Exemplarily, when the calculated training loss value fails to match the preset specified loss threshold, the model parameters in the training model are adjusted to train the entire model, so that the trained decoder converges along with the convergence of the model, where the parameters of the training model include the model parameters of the encoder and the model parameters of the decoder.

[0116] In step 7032, a predictive model is obtained in response to the training loss value between the sample structure data and the predicted structure data reaching a specified loss threshold.

[0117] When the training loss value obtained by the specified loss function reaches the specified loss threshold, the training of the entire training model is determined to be completed, i.e., a prediction model is obtained, where the decoder part in the prediction model corresponds to the specified decoder. In one example, since the predicted structure data to be output in the training process needs to be as close as possible to the input sample structure data, in response to the training loss value between the sample structure data and the predicted structure data being smaller than the specified loss threshold, the training of the training model is determined to be completed, and a prediction model is obtained.

[0118] Preferably, the trained prediction model may be stored in the server and called by the terminal in response to a generation request, i.e., the terminal transmits a generation request to the server, the server calls the prediction model to generate the reconstructed structure data, and returns the generated reconstructed structure data to the terminal. Alternatively, the trained prediction model may be transmitted by the server to the terminal, and the terminal may generate the reconstructed structure data using the prediction model.

[0119] Preferably, in the application process, the reconstructed structure data can be generated by a complete prediction model, for example, the candidate structure data is input to the prediction model, and the prediction model encodes the data to obtain a hidden layer feature representation, and the data is decoded to reconstruct the structure, thereby obtaining reconstructed structure data having a structural similarity relationship with the candidate structure data, and the method can be applied to a scene where it is necessary to generate structure data having a strong similarity to the designated structure. Alternatively, in the application process, the reconstructed structure data can be generated only by a designated decoder in the prediction model. That is, the candidate data of the designated data distribution is sampled to obtain sampling data, and the sampling data is input to the designated decoder, and the designated decoder reconstructs the structure based on the sampling data to obtain the corresponding reconstructed structure data. The method can be applied to an application scene where it is necessary to generate reconstructed structure data whose properties are unknown, and can improve the diversity of the generated structure while ensuring the rationality of the generated structure.

[0120] As described above, the decoder training method according to the embodiment of the present application performs structure reconstruction on the hidden layer feature representation obtained by the decoder to obtain corresponding predicted structure data, and trains the entire training model according to the difference between the predicted structure data and the sample structure data, thereby realizing the decoder training process. Here, the decoder achieves the functions of high-frequency signal reconstruction and noise removal by restoring the high-frequency characteristics compressed and reduced in the hidden layer according to the inverse transformation process of the wavelet transform, and when data is secondarily smoothed due to the reconstruction using the GCN as it is after filtering, there is an accuracy accumulation effect (the final accuracy is proportional to the Nth power of the model prediction accuracy), that is, the prediction accuracy is further reduced in the decoding process after being reduced in the encoding process, which solves the problem that the diversity of generated structure data is low and the generation rate is low, and improves the accuracy of the decoder reconstruction result in the application process.

[0121] For example, when the above decoder training method is applied to a chemical molecule generation scene to obtain new chemical molecules by reconstructing atom types, the reconstruction root mean squared error (RMSE) corresponding to the method of the present application can be reduced by about 10% compared to the method of restoring the high-frequency characteristics compressed and reduced in the hidden layer and directly completing the reconstruction process by GCN, and the reconstruction accuracy can be guaranteed by both the prediction of atom types and graph structure, that is, the stability of the properties of generating new chemical molecules can be greatly improved, and the effectiveness of generating new chemical molecules can be guaranteed.

[0122] 8 shows a method for generating structure data according to an embodiment of the present application, which illustrates the application of the trained prediction model, and in the embodiment of the present application, the trained prediction model is used to complete the generation of reconstructed structure data. The method includes the following steps:

[0123] In step 801, a candidate structure feature representation and a candidate node feature representation of candidate structure data are obtained.

[0124] Exemplarily, the candidate structure data is data waiting for generation of similar structure data, and the candidate structure data is data whose data structure is a graph structure, that is, the candidate structure data is data consisting of at least two nodes and at least one edge. Preferably, the graph structure may be any of the graph structures of an undirected graph, a directed graph, an undirected complete graph, a directed complete graph, etc., and the specific graph structure can be determined based on the data information corresponding to the candidate structure data.

[0125] In one example, the predictive model is used to generate chemical molecules, and the candidate structure data corresponds to the candidate chemical molecules, where the atoms in the molecules are nodes in the graph, and the chemical bonds between the atoms are edges in the graph. Exemplarily, a corresponding candidate molecular graph is generated based on the chemical structure of the candidate chemical molecule, and the candidate molecular graph is used as the candidate structure data, and a candidate molecular structure characteristic representation and a candidate node characteristic representation can be obtained based on the candidate molecular graph, where the candidate molecular structure characteristic representation is an adjacency matrix that records the connection status between the atoms in the candidate chemical molecule, and the candidate node characteristic representation is a matrix that records the atom type that constitutes each atom in the candidate chemical molecule.

[0126] In step 802, a candidate hidden layer representation is generated based on the candidate structural representation and the candidate node representation.

[0127] For example, the candidate structure feature representation and the candidate node feature representation are encoded by a designated encoder in the prediction model to obtain intermediate encoded data, and the candidate hidden layer feature representation is obtained by clustering the intermediate encoded data. In the embodiment of the present application, the intermediate encoded data is data obtained by performing intermediate calculations according to the features encoded by the designated encoder, that is, the data obtained by the intermediate calculations are the mean value and the variance.

[0128] According to the candidate structure feature representation and the candidate node feature representation, obtain the node feature vector of the node of the candidate structure data in the feature space corresponding to at least two frequency bands, obtain the average value data and the variance data between the node feature vectors corresponding to at least two frequency bands, and determine the average value data and the variance data as intermediate coding data. After determining the average value data and the variance data, obtain the hidden layer feature representation by calculating using the reparameterization trick, and the specific determination process is the same as steps 402 to 403. Since this embodiment is a data processing process in the application stage, the description thereof is omitted here.

[0129] In step 803, the candidate hidden layer feature representations are input to a designated decoder for prediction to obtain reconstructed structural data.

[0130] There exists a structure property similarity relationship between the reconstructed structure data and the candidate structure data.

[0131] For example, the candidate hidden layer feature representation is input to a designated decoder, and the designated decoder performs structural reconstruction according to the candidate hidden layer feature representation, so that the reconstructed structural feature representation and the reconstructed node feature representation can be obtained, and the reconstructed structural data can be obtained according to the reconstructed structural feature representation and the reconstructed node feature representation. Specifically, for example, the candidate structural data is a candidate molecular graph corresponding to a candidate chemical molecule, and the reconstructed structural data obtained by prediction is a candidate molecular structure, and the candidate molecular structure and the input candidate chemical molecule have similar chemical properties. When applied correspondingly to the intelligent medicine scene, the candidate molecular structure and the input candidate chemical molecule may have similar pharmacological properties, which supports the research process of substitute drugs and drug optimization.

[0132] As described above, the method for generating structured data according to the embodiments of the present application generates new structured data having a structural property similarity relationship by encoding and reconstructing input candidate structural data using a complete predictive model obtained through training, where the predictive model is a model that employs wavelet encoding and decoding, thereby performing the functions of reconstructing high-frequency signals and removing noise, improving the accuracy of the reconstructed structural data obtained by reconstruction, and ensuring the structural property similarity between the input and output structural data, and can be applied to generating similar structured data.

[0133] 9 shows a method for generating structure data according to an embodiment of the present application, which is a schematic illustration of the application of a trained designated decoder. In the embodiment of the present application, the trained designated decoder is used to complete the generation of reconstructed structure data, and the method includes the following steps:

[0134] In step 901, candidate data for a specified data distribution is obtained.

[0135] In the embodiments of the present application, after the training of the prediction model is completed, the designated decoder in the prediction model is separated and applied as a model for generating reconstructed structure data. The designated data distribution may be a Gaussian distribution, and the candidate data may be custom candidate data input from a terminal, and after the prediction model is trained to convergence, it may be curve data of a structural nature corresponding to a structure generation task, and the structure generation task is a task corresponding to the prediction model, that is, the curve data is data obtained by learning sample structure data in a training set. In some embodiments, the candidate data may be candidate data generated after encoding the input candidate structure data to obtain mean data and variance data.

[0136] In step 902, a preset number of sampled data are obtained by sampling the candidate data.

[0137] Exemplarily, the candidate data is randomly sampled to obtain a preset number of sampling data, where the preset number is the number of reconstruction structure data that needs to be generated by the designated decoder as instructed by the terminal. The sampling data obtained by the sampling is for instructing the hidden layer representation between the nodes and edges in the reconstruction structure data to be generated, where the number of corresponding nodes in each sampling data may be randomly generated or specified, and the number of nodes between the preset number of sampling data may be the same or different. In one example, the sampling data Z is sampled according to Equation 10 to be input to the designated decoder, where N(0,I) represents the candidate data following a normal distribution.

[0138]

number

[0139] In step 903, a preset number of sampling data are input to a designated decoder to obtain a preset number of reconstructed structure data.

[0140] When the sampling data Z obtained by sampling is input to a designated decoder, the designated decoder performs node prediction based on the sampling data, and reconstructs a structural feature representation and a node feature representation corresponding to the reconstructed structural data according to the probability that an edge exists between the nodes, and obtains reconstructed structural data from the structural feature representation and the node feature representation.

[0141] In one example, the designated decoder is used to generate chemical molecules, i.e., the designated decoder generates candidate molecular structures, where the candidate molecular structures are constructed from at least two atomic nodes, and exemplarily, a preset number of sampling data is input to the designated decoder, and the designated decoder obtains a preset number of candidate molecular structures based on the connection relationships between the atomic nodes in the molecular structures learned during training. That is, the designated decoder can generate candidate molecular structures that satisfy chemical rules based on the sampling data. The generated preset number of molecular structures have a certain validity, i.e., the molecular structures are valid under chemical rules, and further support the generation of lead compounds.

[0142] As described above, the method for generating structure data in an embodiment of the present application samples candidate data of a specified data distribution to obtain a certain number of sampled data, and inputs the sampled data into a specified decoder to reconstruct the structure, thereby obtaining reconstructed structure data with a certain degree of effectiveness and improving the diversity of the generated reconstructed structure data.

[0143] For example, the structural data generation method according to the embodiment of the present application is applied to the public dataset ZINC for testing, a designated decoder is obtained by training, and new chemical molecules are generated by the designated decoder. In the process of generating new chemical molecules, 104 random samplings are performed on the normally distributed data N(0,I), and the sampling results are input into the designated decoder obtained by training to obtain 104 newly generated chemical molecules. The obtained new chemical molecules are verified using the open source platform rdkit, and it is found that the new chemical molecules have high uniqueness and novelty under the premise that their validity is guaranteed. Among them, the validity is 98.0%, the uniqueness is 99.1%, and the novelty is 96.5%. The higher the uniqueness and novelty, the higher the diversity of the generated molecules. Overall, the diversity of the generated molecules is improved, and the generation space can be expanded.

[0144] It should be noted that when the above embodiments of the present application are applied to specific products or technologies involving user data (e.g., the method is applied to a recommendation system), the acquisition of such data should obtain the user's permission or consent, and the compound research and the use and processing of the data should all comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0145] 10 is a block diagram showing the structure of a structure data generating device according to one exemplary embodiment of the present application. an acquisition module 1010 for acquiring a structural characteristic representation and a node characteristic representation of sample structure data, the structural characteristic representation being for indicating a connection state between nodes constituting the sample structure data, and the node characteristic representation being for indicating a node type corresponding to the node constituting the sample structure data; an encoding module 1020 for generating hidden layer feature representations based on the structure feature representation and the node feature representation, the hidden layer feature representations for indicating connectivity between nodes in the sample structure data in at least two frequency bands; a decoding module 1030 for inputting the hidden layer feature representation to a trained decoder to perform reconstruction of the structure to obtain predicted structure data; and a training module 1040 for obtaining a designated decoder by training the training-awaiting decoder based on the predicted structure data, the designated decoder performing structure reconstruction on input sampling data to obtain reconstructed structure data, the sampling data being data obtained by sampling candidate data.

[0146] In some alternative embodiments, as shown in FIG. 11, the encoding module 1020: an encoding unit 1021 for encoding the structure feature representation and the node feature representation in the at least two frequency bands respectively to obtain intermediate feature data, the intermediate feature data being for indicating a coupling situation between nodes of the sample structure data in a corresponding frequency band; and a clustering unit 1022 for clustering the intermediate feature data respectively corresponding to the at least two frequency bands based on the specified data distribution to obtain the hidden layer feature representation, wherein the candidate data is data satisfying the specified data distribution.

[0147] In some alternative embodiments, the encoding unit 1021 is further used to input the structural feature representation and the node feature representation into a training-ready encoder, which, when the structural feature representation and the node feature representation are known conditions, determines a connection probability for an i-th prediction node in a prediction node set to establish a connection relationship with each prediction node in the prediction node set, and determines a connection probability distribution corresponding to the i-th prediction node based on the connection probability between the i-th prediction node and each prediction node, and determines the intermediate feature data corresponding to each of the at least two frequency bands based on a fusion result of the connection probability distributions of all prediction nodes in the prediction node set, and the prediction node is for constructing the prediction structural data, where i is a positive integer.

[0148] In some optional embodiments, the encoding unit 1021 is further used to obtain node feature vectors of nodes of the sample structure data in a feature space corresponding to the at least two frequency bands respectively based on the structure feature representation and the node feature representation, obtain average value data and variance data between the node feature vectors corresponding to the at least two frequency bands, and determine the average value data and the variance data as the intermediate feature data.

[0149] In some alternative embodiments, the decoding module 1030: a reconstruction unit 1031 for performing structural reconstruction on the hidden layer representations by the trained decoder to obtain a decoded structural representation and a decoded node representation; and a generating unit 1032 for generating the predicted structure data based on the decoded structure feature representation and the decoded node feature representation.

[0150] In some alternative embodiments, the apparatus comprises: The decoder further includes a training module 1040 for obtaining a training loss value based on a structural difference status between the sample structural data and the predicted structural data, and determining that training of the decoder waiting to be trained is completed and obtaining the designated decoder in response to the training loss value reaching a designated loss threshold, or for iteratively training model parameters of the decoder waiting to be trained in response to failure to match the training loss value with the designated loss threshold.

[0151] In some alternative embodiments, the training module 1040 includes: an obtaining unit 1041 for obtaining distance metric data of the sample structure data and the predicted structure data in a feature space; and a determining unit 1042 for obtaining the training loss value based on the distance metric data and the divergence data.

[0152] The acquisition unit 1041 is further used to acquire divergence data between a node distribution corresponding to the predicted structure data and the specified data distribution, where the node distribution is for indicating the distribution status of the node feature vector of the predicted structure data in the feature space.

[0153] In some alternative embodiments, the acquisition module 1010 is further used to acquire candidate structural feature representations and candidate node feature representations of the candidate structural data; The encoding module 1020 is further used to generate a candidate hidden layer feature representation based on the candidate structural feature representation and the candidate node feature representation; The decoding module 1030 is further used to obtain the reconstructed structure data by inputting the candidate hidden layer feature representation to the designated decoder for prediction, and there is a structural property similarity relationship between the reconstructed structure data and the candidate structure data.

[0154] In some alternative embodiments, the apparatus comprises: The data distribution data storage device further includes a sampling module 1050 for obtaining candidate data of the specified data distribution and sampling the candidate data to obtain a preset number of sampling data.

[0155] The decoding module 1030 is further used for inputting the predetermined number of sampling data to the designated decoder to obtain the predetermined number of the reconstructed structure data.

[0156] In some alternative embodiments, the designation decoder is adapted to generate a candidate molecular structure comprised of at least two atomic nodes; The decoding module 1030 is further used to input the predetermined number of sampling data to the designated decoder, and obtain the predetermined number of candidate molecular structures based on the connection relationships between atomic nodes in the molecular structures learned during training by the designated decoder.

[0157] In some alternative embodiments, when the trained designated decoder is used to generate a molecular structure, the acquired model 1010 is further used to acquire a sample chemical molecule, which is a known molecule that satisfies an atomic bonding criterion and is composed of at least two atoms, convert the sample chemical molecule into a sample molecular graph, the data structure of which is a graph structure, wherein nodes of the sample molecular graph are for representing the at least two atoms in the sample chemical molecule, and edges in the sample molecular graph are for representing chemical bonds between atoms in the sample chemical molecule, determine an adjacency matrix corresponding to the sample molecular graph as the structure feature representation, and determine a node matrix corresponding to the sample molecular graph as the node feature representation.

[0158] As described above, the structure data generation device according to an embodiment of the present application obtains a hidden layer feature representation using a structure feature representation and a node feature representation corresponding to sample structure data, and then iteratively trains a decoder waiting to be trained using the hidden layer feature representation to obtain a designated decoder, which can generate structure data using input sampling data. In other words, as necessary, a variety of structure data can be quickly generated by the designated decoder obtained through training, thereby improving the efficiency of generation of structure data and the diversity of its generation.

[0159] In addition, the structure data generating device according to the above embodiment has been described by taking the division of each of the above functional modules as an example, but in actual application, the above functions can be assigned to different functional modules to be completed as necessary, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the above functions. In addition, the structure data generating device according to the above embodiment belongs to the same concept as the embodiment of the structure data generating method, and the specific implementation process thereof can be referred to the embodiment of the method, so the description will be omitted here.

[0160] 12 is a diagram showing the structure of a server according to an exemplary embodiment of the present application. Specifically, it includes the following structures:

[0161] The server 1200 includes a CPU (Central Processing Unit) 1201, a system memory 1204 including a RAM (Random Access Memory) 1202 and a ROM (Read Only Memory) 1203, and a system bus 1205 connecting the system memory 1204 and the CPU 1201. The server 1200 further includes a mass storage device 1206 for storing an operating system 1213, application programs 1214 and other program modules 1215.

[0162] The mass storage device 1206 is connected to the CPU 1201 through a mass storage controller (not shown) that is connected to the system bus 1205. The mass storage device 1206 and its associated computer readable media provide non-volatile storage for the server 1200.

[0163] Without loss of generality, computer-readable media may include computer storage media and communication media. The system memory 1204 and mass storage device 1206 may be collectively referred to as memory.

[0164] According to various embodiments of the present application, the server 1200 may be connected to a network 1212 by a network interface unit 1211 connected to the system bus 1205, or may be connected to other types of networks or remote computer systems (not shown) by using the network interface unit 1211.

[0165] The memory further includes one or more programs stored in the memory and configured to be executed by the CPU.

[0166] An embodiment of the present application further provides a computing device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set that is loaded and executed by the processor to realize the method for generating structure data according to each of the above method embodiments. Preferably, the computing device may be a terminal or a server.

[0167] An embodiment of the present application further provides a computer-readable storage medium having stored thereon at least one instruction, at least one program, code set or instruction set that, when loaded and executed by a processor, realizes the method of generating structure data relating to each of the above method embodiments.

[0168] An embodiment of the present application further provides a computer program product or a computer program comprising computer instructions stored in a computer readable storage medium, the computer instructions being read from the computer readable storage medium by a processor of a computing device and executed to cause the computing device to perform the method for generating structure data according to any of the above embodiments.

[0169] Preferably, the computer-readable recording medium may include a ROM (Read Only Memory), a RAM (Random Access Memory), an SSD (Solid State Drives), an optical disk, etc. Here, the RAM may include a ReRAM (Resistance Random Access Memory) and a DRAM (Dynamic Random Access Memory). The sequence numbers of the above embodiments of the present application are merely for explanation and do not indicate the superiority or inferiority of the embodiments.

Claims

1. 1. A method for generating structure data executed by a computing device, comprising: a step of acquiring a structural characteristic representation and a node characteristic representation of sample structure data, the structural characteristic representation being for indicating a connection state between nodes constituting the sample structure data, and the node characteristic representation being for indicating a node type corresponding to a node constituting the sample structure data; generating hidden layer feature representations based on the structural feature representation and the node feature representation, the hidden layer feature representations being for indicating connectivity between nodes in the sample structural data in at least two frequency bands; obtaining predicted structure data by inputting the hidden layer feature representation to a trained decoder to perform structure reconstruction; A step of obtaining a trained decoder by training the trained decoder based on the predicted structure data, the trained decoder being for performing structure reconstruction on input sampling data to obtain reconstructed structure data, the sampling data being data obtained by sampling candidate data; How structure data is generated.

2. generating a hidden layer representation based on the structural representation and the node representation, obtaining intermediate feature data by encoding the at least two frequency bands based on the structure feature representation and the node feature representation, respectively, the intermediate feature data being for indicating a coupling situation between nodes of the sample structure data in a corresponding frequency band; obtaining the hidden layer feature representation by clustering the intermediate feature data respectively corresponding to the at least two frequency bands based on a specified data distribution, wherein the candidate data is data satisfying the specified data distribution; The method of claim 1.

3. said step of obtaining intermediate feature data by encoding the at least two frequency bands based on the structural feature representation and the node feature representation, respectively, comprising: inputting the structural feature representation and the node feature representation into a training-ready encoder, in which, when the structural feature representation and the node feature representation are known conditions, the training-ready encoder determines a connection probability for an i-th prediction node in a prediction node set to establish a connection relationship with each prediction node in the prediction node set, and determines a connection probability distribution corresponding to the i-th prediction node based on the connection probability between the i-th prediction node and each prediction node, and determines the intermediate feature data respectively corresponding to the at least two frequency bands based on a fusion result of the connection probability distributions of all prediction nodes in the prediction node set, wherein the prediction node is for constructing the prediction structural data, and i is a positive integer; The method of claim 2.

4. said step of obtaining intermediate feature data by encoding the at least two frequency bands based on the structural feature representation and the node feature representation, respectively, comprising: obtaining node feature vectors in a feature space corresponding to the at least two frequency bands of the nodes of the sample structure data based on the structure feature representation and the node feature representation; obtaining mean and variance data between the node feature vectors corresponding to the at least two frequency bands; determining the mean value data and the variance data as the intermediate feature data; The method of claim 2.

5. The step of obtaining predicted structure data by inputting the hidden layer feature representation to a trained decoder to perform structure reconstruction, performing structural reconstruction on the hidden layer representations by the trained decoder to obtain a decoded structural representation and a decoded node representation; generating the predicted structure data based on the decoded structure feature representation and the decoded node feature representation; 5. The method according to any one of claims 1 to 4.

6. said step of obtaining a designated decoder by training said waiting-to-train decoder based on said prediction structure data comprises: obtaining a training loss value based on a structural difference situation between the sample structural data and the predicted structural data; in response to the training loss value reaching a designated loss threshold, determining that training of the decoder waiting to be trained is complete to obtain the designated decoder; or and in response to a failure to match the training loss value with the specified loss threshold, iteratively training model parameters of the waiting-to-train decoder.

5. The method according to any one of claims 1 to 4.

7. The step of obtaining a training loss value based on a structural difference situation between the sample structural data and the predicted structural data includes: obtaining distance metric data for the sample structure data and the predicted structure data in a feature space; obtaining divergence data between a node distribution corresponding to the predicted structure data and the specified data distribution, the node distribution indicating a distribution state of a node feature vector of the predicted structure data in the feature space, and the divergence data indicating a degree of difference between the node distribution and the specified data distribution; and obtaining the training loss value based on the distance metric data and the divergence data. The method according to claim 6.

8. obtaining a candidate structure representation and a candidate node representation of the candidate structure data; generating a candidate hidden layer representation based on the candidate structural representation and the candidate node representation; and obtaining the reconstructed structure data by inputting the candidate hidden layer feature representation to the designated decoder for prediction, wherein there is a structure property similarity relationship between the reconstructed structure data and the candidate structure data.

5. The method according to any one of claims 1 to 4.

9. obtaining candidate data for a specified data distribution; A step of sampling the candidate data to obtain a preset number of sampling data; inputting the predetermined number of sampling data to the designated decoder to obtain the predetermined number of reconstructed structure data; 5. The method according to any one of claims 1 to 4.

10. The designated decoder is used to generate a candidate molecular structure composed of at least two atomic nodes; The step of inputting the predetermined number of sampling data to the designated decoder to obtain the predetermined number of reconstructed structure data includes: inputting the predetermined number of sampling data to the designated decoder, and obtaining the predetermined number of candidate molecular structures based on connection relationships between atomic nodes in molecular structures learned by the designated decoder during training; 10. The method of claim 9.

11. If the trained designated decoder is to be used to generate molecular structures, prior to the step of obtaining structural and node feature representations of the sample structure data, obtaining a sample chemical molecule, the sample chemical molecule being a known molecule that satisfies an atomic bonding criterion and is composed of at least two atoms; converting the sample chemical molecule into a sample molecular graph, the data structure of which is a graph structure, wherein a node of the sample molecular graph is for representing the at least two atoms in the sample chemical molecule, and an edge in the sample molecular graph is for representing a chemical bond between atoms in the sample chemical molecule; determining an adjacency matrix corresponding to the sample molecular graph as the structural feature representation; determining a node matrix corresponding to the sample molecular graph as the node feature representation; 5. The method according to any one of claims 1 to 4.

12. an acquisition module for acquiring a structural characteristic representation and a node characteristic representation of sample structure data, the structural characteristic representation indicating a connection state between nodes constituting the sample structure data, and the node characteristic representation indicating a node type corresponding to the node constituting the sample structure data; an encoding module for generating hidden layer feature representations based on the structural feature representation and the node feature representation, the hidden layer feature representations for indicating connectivity between nodes in the sample structure data in at least two frequency bands; a decoding module for inputting the hidden layer feature representation to a trained decoder to perform structure reconstruction to obtain predicted structure data; a training module for obtaining a designated decoder by training the training-awaiting decoder based on the predicted structure data, the designated decoder reconstructing a structure of input sampling data to obtain reconstructed structure data, the sampling data being data obtained by sampling candidate data; A structured data generator.

13. A processor and a memory are included. The memory stores at least one instruction, at least one program, code set or instruction set that is loaded and executed by the processor to implement the method for generating structure data according to any one of claims 1 to 11. Computer device.

14. At least one program code is stored, which is loaded and executed by a processor to realize the method for generating structure data according to any one of claims 1 to 11. A computer readable storage medium.

15. A computer program or instructions which, when executed by a processor, implements the method for generating structure data according to any one of claims 1 to 11, Computer program products.

Citation Information

Patent Citations

  • Substance structure analysis device, method and program

    JP2020139914A

  • Chemical compound structure automatic creation device for automatically creating chemical compound structure, chemical compound structure automatic creation system and chemical compound structure automatic creation method

    JP2021081769A

  • Target-to-catalyst translation networks

    WO2021081390A1