Method and device for model training and product prediction

Abstract

The embodiment of the present application provides a method and device for model training and product prediction, the training method comprising: obtaining a first reactant and a first product of the first reactant; performing noise processing on the first product to obtain the first product under different noise values; inputting the first reactant and the first product under different noise values ​​into a prediction model to obtain gradient field information of the adjacency matrix of the first product output by the prediction model under different noise values; and jointly training the prediction model based on the gradient field information of the adjacency matrix of the first product under different noise values ​​to obtain a trained prediction model. That is, the present application jointly trains the prediction model by using the first reactant and the first product under different noise values, and the obtained prediction model is used to predict the gradient field information of the adjacency matrix of the product of the second reactant, which can achieve prediction of products of different types of reactants.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technology, and in particular to a method and apparatus for model training and product prediction. Background Art

[0002] In the fields of chemistry and pharmaceuticals, predicting products based on reactants is of great significance.

[0003] However, there are many types of organic chemical reactions, and with the development of chemical research technology, new chemical reactions emerge in an endless stream. How to accurately predict the products of reactants has become a technical problem that needs to be solved urgently in this field. Summary of the Invention

[0004] The embodiments of the present application provide a method and apparatus for model training and product prediction, which predicts the product of any reactant through a prediction model, thereby achieving accurate prediction of the products of different reactants.

[0005] In a first aspect, an embodiment of the present application provides a model training method, comprising:

[0006] obtaining a first reactant and a first product of the first reactant;

[0007] performing noise processing on the first product to obtain first products with different noise values;

[0008] Inputting the first reactant and the first product under different noise values ​​into a prediction model, and obtaining gradient field information of the adjacency matrix of the first product under different noise values ​​output by the prediction model;

[0009] The prediction model is trained based on the gradient field information of the adjacency matrix of the first product under different noise values ​​to obtain the trained prediction model, wherein the trained prediction model is used to predict the gradient field information of the adjacency matrix of the product of the second reactant, and the gradient field information is used to indicate the changing trend of the generation probability of the product of the second reactant.

[0010] In some embodiments, the first feature extraction subunit and the second feature extraction subunit are both encoders in the converter, and the feature decoding unit is a decoder in the converter.

[0011] In some embodiments, the step of adding noise to the adjacency matrix of the first product to obtain the adjacency matrix of the first product under different noise values ​​includes:

[0012] A Gaussian distribution is used to add noise, and different noise values ​​are added to the adjacency matrix of the first product to obtain the adjacency matrix of the first product under different noise values.

[0013] In some embodiments, the p-th molecular feature extraction unit is a multi-layer perceptron MLP.

[0014] In some embodiments, the correspondence between the noise value and the model parameter is the correspondence between the noise value and the parameter in the MLP.

[0015] In a second aspect, an embodiment of the present application provides a method for predicting a product, comprising:

[0016] Obtaining a second reactant to be predicted and K preset noise values, where K is a positive integer less than or equal to L;

[0017] For an i-th noise value among the K noise values, determining a prediction model corresponding to the i-th noise value, where the prediction model is trained using the training method described in the first aspect, and i is a positive integer from 1 to K;

[0018] According to the target adjacency matrix of the second reactant and the second product at the (i-1)th noise value, sampling is performed in the gradient field information predicted by the prediction model corresponding to the (i)th noise value to obtain the target adjacency matrix of the second product at the (i)th noise value, where the second product is the product of the second reactant;

[0019] The second product is determined according to a target adjacency matrix of the second product under a K-th noise value, wherein the K-th noise value is a minimum value among the K noise values.

[0020] In some embodiments, the encoding module includes a first encoding submodule and a second encoding submodule;

[0021] The first encoding submodule is used to process the adjacency matrix and the node feature matrix of the second reactant to obtain first feature information of the second reactant;

[0022] The second encoding submodule is used to process the t-1th adjacency matrix of the second product under the i-th noise value to obtain first feature information of the t-1th adjacency matrix of the second product.

[0023] In some embodiments, the first encoding submodule includes a first atomic feature extraction unit and a first molecular feature extraction unit;

[0024] The first atomic feature extraction unit is used to process the adjacency matrix and the node feature matrix of the second reactant to obtain an embedded representation of each atom in the second reactant;

[0025] The first molecular feature extraction unit is used to perform feature interaction on the embedded representation of each atom in the second reactant to obtain first feature information of the second reactant.

[0026] In some embodiments, the second encoding submodule includes a second atomic feature extraction unit and a second molecular feature extraction unit;

[0027] The second atom feature extraction unit is used to process the t-1th adjacency matrix of the second product under the i-th noise value to obtain an embedded representation of each atom corresponding to the t-1th adjacency matrix of the second product;

[0028] The second molecular feature extraction unit is used to perform feature interaction on the embedded representation of each atom corresponding to the t-1th adjacency matrix of the second product to obtain first feature information of the t-1th adjacency matrix of the second product.

[0029] In some embodiments, the decoding module includes a feature extraction unit and a feature decoding unit;

[0030] The feature extraction unit is used to obtain second characteristic information of the second reactant and second feature information of the t-1th adjacency matrix of the second product according to the first feature information of the second reactant and the first feature information of the t-1th adjacency matrix of the second product;

[0031] The feature decoding unit is used to obtain the t-1th gradient field information of the adjacency matrix of the second product under the i-th noise value based on the second characteristic information of the second reactant and the second feature information of the t-1th adjacency matrix of the second product.

[0032] In some embodiments, the feature extraction unit includes a first feature extraction subunit and a second feature extraction subunit;

[0033] The first feature extraction subunit is used to obtain second characteristic information of the second reactant based on the first feature information of the second reactant;

[0034] The second feature extraction subunit is configured to obtain second feature information of the t-1th adjacency matrix of the second product based on the first feature information of the t-1th adjacency matrix of the second product.

[0035] In some embodiments, the first feature extraction subunit and the second feature extraction subunit are both encoders in the converter, and the feature decoding unit is a decoder in the converter.

[0036] In some embodiments, if i is 1, each element in the target adjacency matrix of the second product under the (i-1)th noise value conforms to a first normal distribution.

[0037] Optionally, the variance of the first normal distribution is a positive number less than or equal to 3.

[0038] In a third aspect, an embodiment of the present application provides a model training device, comprising:

[0039] an acquisition unit, configured to acquire a first reactant and a first product of the first reactant;

[0040] a noise adding unit, configured to perform noise adding processing on the first product to obtain first products with different noise values;

[0041] A prediction unit, configured to input the first reactant and the first product under different noise values ​​into a prediction model, and obtain gradient field information of an adjacency matrix of the first product under different noise values ​​output by the prediction model;

[0042] A training unit is used to train the prediction model based on the gradient field information of the adjacency matrix of the first product under different noise values ​​to obtain the trained prediction model, wherein the trained prediction model is used to predict the gradient field information of the adjacency matrix of the product of the second reactant, and the gradient field information is used to indicate the changing trend of the generation probability of the product of the second reactant.

[0043] In a fourth aspect, an embodiment of the present application provides a device for predicting a product, comprising:

[0044] an acquisition unit, configured to obtain a second reactant and K preset noise values;

[0045] a determining unit, configured to determine, for an i-th noise value among the K noise values, a prediction model corresponding to the i-th noise value, wherein the prediction model is obtained by training using the training method described in the first aspect, and i is a positive integer from 1 to K;

[0046] a sampling unit, configured to sample, based on the target adjacency matrix of the second reactant and the second product at the (i-1)th noise value, the gradient field information predicted by the prediction model corresponding to the (i)th noise value, to obtain the target adjacency matrix of the second product at the (i)th noise value, where the second product is a product of the second reactant;

[0047] A prediction unit is configured to determine the second product according to a target adjacency matrix of the second product under a K-th noise value, wherein the K-th noise value is a minimum value among the K noise values.

[0048] In a fifth aspect, an embodiment of the present application provides a computing device, including a processor and a memory;

[0049] The memory is used to store computer programs;

[0050] The processor is configured to execute the computer program to implement the method described in any one of the first to second aspects above.

[0051] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, which includes computer instructions. When the instructions are executed by a computer, the computer implements the method described in any one of the first to second aspects.

[0052] In the seventh aspect, an embodiment of the present application provides a computer program product, which includes a computer program, and the computer program is stored in a readable storage medium. At least one processor of a computer can read the computer program from the readable storage medium, and the at least one processor executes the computer program so that the computer implements the method described in any one of the first to second aspects.

[0053] The model training and product prediction method and device provided by the embodiment of the present application, by obtaining a first reactant and a first product of the first reactant; Noise processing is performed on the first product to obtain the first product under different noise values; The first reactant and the first product under different noise values ​​are input into the prediction model to obtain the gradient field information of the adjacency matrix of the first product output by the prediction model under different noise values; According to the gradient field information of the adjacency matrix of the first product under different noise values, the prediction model is jointly trained to obtain the trained prediction model. That is, the embodiment of the present application is jointly trained by the first reactant and the first product under different noise values ​​to the prediction model, and the prediction model under different noise values ​​can be obtained, the prediction model is used to predict the gradient field information of the adjacency matrix of the product of the second reactant, and the gradient field information is used to indicate the change trend of the generation probability of the product of the second reactant, so that when the product is predicted, the product can be accurately predicted according to the gradient field information of the adjacency matrix of the product. In addition, the prediction model of the present application does not limit the type of reactant, and thus can achieve the prediction of the products of different types of reactants. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0055] Figure 1 A schematic diagram of a system architecture involved in an embodiment of the present application;

[0056] Figure 2 A flowchart of a model training method provided in one embodiment of the present application;

[0057] Figure 3A A training diagram of a prediction model involved in an embodiment of the present application;

[0058] Figure 3B A schematic diagram of the structure of the prediction model involved in the embodiment of the present application;

[0059] Figure 3C A schematic diagram of the structure of the prediction model involved in the embodiment of the present application;

[0060] Figure 3D A schematic diagram of the structure of the prediction model involved in the embodiment of the present application;

[0061] Figure 3E A schematic diagram of the structure of the prediction model involved in the embodiment of the present application;

[0062] Figure 3F A schematic diagram of the structure of the prediction model involved in the embodiment of the present application;

[0063] Figure 3G A schematic diagram of the structure of the prediction model involved in the embodiment of the present application;

[0064] Figure 3H A schematic diagram of the structure of the prediction model involved in the embodiment of the present application;

[0065] Figure 4 Schematic diagram of the training process of the prediction model involved in the embodiment of the present application;

[0066] Figure 5 A schematic flow chart of a method for predicting a product provided in one embodiment of the present application;

[0067] Figure 6A A prediction diagram of a prediction model involved in an embodiment of the present application;

[0068] Figure 6B Another prediction diagram of the prediction model involved in the embodiment of the present application;

[0069] Figure 6C Another prediction diagram of the prediction model involved in the embodiment of the present application;

[0070] Figure 6D Another prediction diagram of the prediction model involved in the embodiment of the present application;

[0071] Figure 6E Another prediction diagram of the prediction model involved in the embodiment of the present application;

[0072] Figure 6F A schematic diagram of the prediction process of the prediction model involved in the embodiment of the present application;

[0073] Figure 7 A schematic diagram of the structure of a training device for a prediction model provided in an embodiment of the present application;

[0074] Figure 8 A schematic diagram of a structure of a prediction device for a product provided in an embodiment of the present application;

[0075] Figure 9 This is a block diagram of a computing device involved in an embodiment of the present application. DETAILED DESCRIPTION

[0076] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0077] It should be understood that in the embodiments of the present invention, "B corresponding to A" means that B is associated with A. In one implementation, B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B based solely on A; B can also be determined based on A and / or other information.

[0078] In the description of the present application, unless otherwise specified, "plurality" means two or more than two.

[0079] In addition, to facilitate the clear description of the technical solutions of the embodiments of the present application, in the embodiments of the present application, the words "first" and "second" are used to distinguish between identical or similar items with substantially the same functions and effects. Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or execution order, and the words "first" and "second" do not necessarily mean different.

[0080] The embodiments of the present application are applied to the field of software testing technology, and in particular to checking the legality of demand data, so as to stably and efficiently generate test cases based on the legal demand data.

[0081] To facilitate understanding of the embodiments of the present application, the following briefly introduces the relevant concepts involved in the embodiments of the present application:

[0082] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0083] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0084] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0085] Natural language processing (NLP) is a key area of ​​research in computer science and artificial intelligence. It studies the theories and methods that enable effective communication between humans and computers using natural language. Natural language processing (NLP) integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language we use in everyday life—and is closely linked to the study of linguistics. Natural language processing technologies typically include text processing, semantic understanding, machine translation, robotic question answering, and knowledge graphs.

[0086] Natural language models: These models use statistical models to convert large amounts of human language into machine language for cognition, understanding, and generation. Specific applications include machine translation and automated question-answering.

[0087] The biggest advantage of AI technology is that it can digest a large amount of learning data through the self-learning process in a short period of time, achieving the goal of self-learning.

[0088] Based on this, the embodiments of the present application utilize AI technology to predict the products of reactants. Specifically, a large amount of organic chemical reaction data is used to train the prediction model. For example, the prediction model is trained using a first reactant and the first product of the first reactant. This allows the prediction model to learn the hidden laws of the reactants. In later predictions, the products of the reactants can be accurately predicted. For example, it can predict products that the reaction template cannot predict, or predict unknown products, greatly improving the research and development efficiency of drugs, etc.

[0089] The application scenarios of this application include but are not limited to medical, biological, scientific research and other fields, such as drug production, drug research and development, vaccine research and development, etc.

[0090] In some embodiments, the system architecture of the embodiment of the present application is as follows Figure 1 shown.

[0091] Figure 1 This is a schematic diagram of a system architecture involved in an embodiment of the present application, including a user device 101, a data acquisition device 102, a training device 103, an execution device 104, a database 105 and a content library 106.

[0092] The data acquisition device 102 is used to read training data from the content library 106 and store the read training data in the database 105. The training data involved in the embodiment of the present application includes multiple reactants and products used for model training. For ease of description, the reactant used for model training is referred to as the first reactant in this application, and the product of the first reactant is known and is recorded as the first product.

[0093] In some embodiments, the user device 101 is used to perform labeling operations on the training data in the database 105 .

[0094] The training device 103 trains the prediction model based on the training data maintained in the database 105 so that the trained prediction model can accurately predict the product of the reactants. The prediction model obtained by the training device 103 can be applied to different systems or devices.

[0095] In one possible implementation, Figure 1In the embodiment, the execution device 104 is configured with an I / O interface 107 for data exchange with external devices. For example, the I / O interface receives the reactant to be predicted (e.g., the second reactant) sent by the user device 101. The calculation module 109 in the execution device 104 uses the trained prediction model to process the input second reactant, outputs the gradient field information of the adjacency matrix of the product of the second reactant, determines the product of the second reactant based on the gradient field information, and sends the product of the second reactant to the user device 101 via the I / O interface.

[0096] The user device 101 may include a mobile phone, a tablet computer, a laptop computer, a PDA, a mobile internet device (MID), or other terminal devices with a browser installed.

[0097] The execution device 104 may be a server.

[0098] For example, the server may be a computing device such as a rack server, a blade server, a tower server, or a cabinet server, etc. The server may be an independent test server or a test server cluster composed of multiple test servers.

[0099] In this embodiment, the execution device 104 is connected to the user device 101 via a network. The network can be a wireless or wired network such as an intranet, the Internet, the Global System of Mobile Communication (GSM), Wideband Code Division Multiple Access (WCDMA), a 4G network, a 5G network, Bluetooth, Wi-Fi, or a call network.

[0100] It should be noted that the attached Figure 1 This is merely a schematic diagram of a system architecture provided by an embodiment of the present application. The positional relationships between the devices, components, modules, etc. shown in the diagram do not constitute any limitation. In some embodiments, the data acquisition device 102, user device 101, training device 103, and execution device 104 can be the same device. The database 105 can be distributed on a single server or multiple servers, and the content library 106 can be distributed on a single server or multiple servers.

[0101] The following describes the technical solutions of the embodiments of the present application in detail through some embodiments. The following embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0102] First combine Figure 2 The training process of the prediction model involved in the embodiments of the present application is introduced.

[0103] Figure 2 A flow chart of a model training method provided in one embodiment of the present application is shown in FIG. Figure 2 Shown, including:

[0104] S201: Obtain a first reactant and a first product of the first reactant.

[0105] The execution subject of the embodiment of the present application is a device with a model training function, such as a model training device, which can be a computing device, or a part of a computing device, such as a processor in a computing device. For example, the above-mentioned model training device can be Figure 1 Training equipment in Figure 1 The training device in the present invention can be understood as a computing device, or a processor in a computing device, etc.

[0106] For ease of description, the following embodiments are introduced by taking the execution subject as a computing device as an example.

[0107] In one example, multiple reactants with known products are collected from the OAS database (Observed Antibody Space database) for model training. For ease of description, the reactant used for model training is denoted as the first reactant, and the product of the first reactant is denoted as the first product.

[0108] It should be noted that the model training of the embodiment of the present application is an iterative process, that is, the model is trained for the first time using the first first reactant and the corresponding first product, and then the model after the first training is trained for the second time using the second first reactant and the corresponding first product, and so on, until the model training is completed.

[0109] The process of training the model using each first reactant and the product of the first reactant is the same. For ease of description, this application uses the training process of the model using a first reactant and the corresponding first biological substance as an example to illustrate.

[0110] In some embodiments, using (G r ,G p ) represents a chemical reaction, G r Represents the reactant, G p Represents the product. r =(V,A r ,Z r), V represents the atomic set of the reactant, that is, the types of atoms included in the reactant. For example, if the reactant is H2O, the atomic set of the reactant includes two atoms, H (hydrogen) and O (oxygen). The size of the atomic set is equal to the number of atoms |V| = N. r ∈R N×N represents the adjacency matrix, Z r Represents the atomic characteristic matrix of the reactants. For example, the atomic characteristic matrix of the reactants includes information such as the type of each atom in the reactants, the chemical properties of the atoms (such as whether they are aromatic), and the charge of the atoms. p =(V,a p ,Z p ) is defined similarly to the reactants.

[0111] The goal of organic chemical reaction prediction is to give a reactant G r Predicted product G p Because the atoms in reactants and products in organic chemical reactions correspond one-to-one, the only change before and after the reaction is the chemical bond connectivity. From the perspective of molecular graph representation, this is the change in the adjacency matrix before and after the chemical reaction. In other words, the change before and after the reaction is mainly the adjacency matrix A, and the atomic set does not change. Therefore, predicting the product only requires predicting its corresponding adjacency matrix, and the entire product structure can be restored based on this adjacency matrix.

[0112] It should be noted that the adjacency matrix of this application may be different from the traditional adjacency matrix, for example, r and A p Each value in is not necessarily just 0 or 1 (0 represents no connection between the corresponding atomic pairs, 1 represents a connection between the corresponding atomic pairs), but can also be multiple discrete or continuous values, such as 0, 1, 2, and 3, which are used to represent the four connection situations of no connection, single bond connection, double bond connection, and triple bond connection, respectively.

[0113] S202: performing noise processing on the first product to obtain first products with different noise values.

[0114] S203 , inputting the first reactant and the first product under different noise values ​​into the prediction model, and obtaining the gradient field information of the adjacency matrix of the first product under different noise values ​​output by the prediction model.

[0115] The gradient field information is used to indicate a changing trend of the generation probability of the first product.

[0116] like Figure 3AAs shown, after the first product is subjected to noise processing, the first reactant and the first product under different noise values ​​are respectively input into the prediction model to obtain the gradient field information of the adjacency matrix of the first product output by the prediction model corresponding to different noise values. The gradient field information is used to represent the changing trend of the generation probability of the first product.

[0117] From the above description, we can see that the first product is G p =(V,A p ,Z p ) represents, where V represents the set of atoms and the set size is the number of atoms |V| = N, A p ∈R N×N Represents the adjacency matrix of the first product, Z p Atomic feature matrix representing the first product.

[0118] In some embodiments, the noise processing of the first product in S202 may be performed on V, A p 、Z p At least one of them is subjected to noise addition processing.

[0119] In some embodiments, the noise addition process of the first product in S202 includes: obtaining an adjacency matrix of the first product; performing noise addition process on the adjacency matrix of the first product to obtain an adjacency matrix of the first product under different noise values. p Perform noise processing.

[0120] This application does not limit the specific method of adding noise to the adjacency matrix of the first product involved above.

[0121] In a possible implementation, a Gaussian distribution is used to add noise, and different noise values ​​are added to the adjacency matrix of the first product to obtain the adjacency matrix of the first product under different noise values.

[0122] For example, different noise values ​​are added to the adjacency matrix of the first product according to the following formula (1), to obtain the adjacency matrix of the first product under different noise values:

[0123]

[0124] in, It is the adjacency matrix of the first product obtained after adding noise, and the noise distribution is obtained through the Gaussian distribution N(A p ,σ) is obtained. μ, σ are two parameters of Gaussian distribution, μ controls the mean of the distribution, and σ controls the standard deviation of the distribution. Among them, the adjacency matrix A of the first product is p as the μ parameter of the Gaussian distribution.

[0125] Since the adjacency matrix of the first product is a triangular symmetric matrix, the above formula (1) means that only the upper triangular matrix is ​​perturbed by noise. In order to maintain the symmetry of the adjacency matrix of the first product, the lower triangular matrix can be obtained symmetrically.

[0126] By adding noise as above, the adjacency matrix A of the first product is p Add different levels of noise (This is done by adding noise using a Gaussian distribution with varying standard deviations σ). This yields L sets of noisy data, each consisting of the adjacency matrix of the first reactant and the adjacency matrix of the first product under a given noise value. These L sets of data are then used to jointly train the prediction model.

[0127] After the adjacency matrix of the first product is subjected to noise processing through the above steps, the corresponding S203 includes S203-A and S203-B:

[0128] S203-A. Obtain the adjacency matrix and node feature matrix of the first reactant, and the node feature matrix of the first product.

[0129] S203-B. Input the adjacency matrix and node feature matrix of the first reactant, the node feature matrix of the first product and the adjacency matrix under different noise values ​​into the prediction model to obtain the gradient field information of the adjacency matrix of the first product output by the prediction model corresponding to different noise values.

[0130] The prediction model of the embodiment of the present application does not directly fit the adjacency matrix of the generated object, but fits the changing trend of the generation probability of the generated object, that is, the gradient field information of the adjacency matrix of the generated object, which is expressed as In this way, in the subsequent actual product prediction process, the adjacency matrix of the product predicted by the prediction model can be optimized in the direction of the gradient field information to obtain the adjacency matrix of the product. That is, by sliding along the gradient field information direction of the adjacency matrix of the product, the adjacency matrix of the product will get closer and closer to the high probability density area, and finally generate an adjacency matrix of the product with better quality.

[0131] This application does not limit the specific network structure of the prediction model. For example, the prediction model can be a graph attention network, a converter, a variational autoencoder, etc.

[0132] In some embodiments, as Figure 3B As shown, the prediction model includes an encoding module and a decoding module. In this case, the above S203-B includes:

[0133] S203-B1. For the i-th noise value among different noise values, the adjacency matrix and atomic characteristic matrix of the first reactant, as well as the atomic characteristic matrix and the adjacency matrix of the first product under the i-th noise value are input into the encoding module to obtain the first characteristic information of the first reactant and the first characteristic information of the first product under the i-th noise value output by the encoding module.

[0134] S203-B2, inputting the first characteristic information of the first reactant and the first characteristic information of the first product under the i-th noise value into a decoding module, and obtaining the gradient field information of the first generated adjacency matrix under the i-th noise value output by the decoding module.

[0135] The following introduces the model training process in combination with the network structure of the encoding module.

[0136] In some embodiments, as Figure 3C As shown, the encoding module includes a first encoding submodule and a second encoding submodule. In this case, the above S203-B1 includes:

[0137] S203-B11. Input the adjacency matrix and node feature matrix of the first reactant into the first encoding submodule to obtain first feature information of the first reactant.

[0138] S203-B12: Input the atomic characteristic matrix of the first product and the adjacency matrix under the i-th noise value into the second encoding submodule to obtain first characteristic information of the first product under the i-th noise value.

[0139] The embodiment of the present application does not limit the network structure of the first encoding submodule and the second encoding submodule. Optionally, the network structures of the first encoding submodule and the second encoding submodule may be different.

[0140] In some embodiments, the network structure of the first encoding submodule and the second encoding submodule is the same. For ease of description, the first encoding submodule and the second encoding submodule are replaced by the pth encoding submodule. For example, when p is 1, the pth encoding submodule is the first encoding submodule, and when p is 2, the pth encoding submodule is the second encoding submodule.

[0141] The p-th encoding submodule includes a p-th atomic feature extraction unit and a p-th molecular feature extraction unit. In this case, the above S203-B1 or S203-B2 can be implemented by the following steps:

[0142] S203-B01, inputting the adjacency matrix and node feature matrix of the target object into the p-th atom feature extraction unit to obtain an embedded representation of each atom in the target object;

[0143] S203-B02, inputting the embedded representation of each atom in the target object into the p-th molecular feature extraction unit for feature interaction to obtain first feature information of the target object;

[0144] Among them, when p is 1, the pth encoding submodule is the first encoding submodule, and the adjacency matrix and node feature matrix of the target object are the adjacency matrix and node feature matrix of the first reactant; when p is 2, the pth encoding submodule is the second encoding submodule, and the adjacency matrix and node feature matrix of the target object are the atomic property matrix of the first product and the adjacency matrix of the first product under the i-th noise value.

[0145] For example, Figure 3D As shown, the first encoding submodule includes a first atomic feature extraction unit and a first molecular feature extraction unit, and the above S203-B1 includes: inputting the adjacency matrix and the node feature matrix of the first reactant into the first atomic feature extraction unit for atomic feature extraction to obtain an embedding representation of each atom in the first reactant; then, inputting the embedding representation of each atom in the first reactant into the first molecular feature extraction unit for atomic feature interaction to obtain first feature information of the first reactant, wherein the first feature information of the first reactant can be understood as molecular structure characteristic information of the first reactant.

[0146] Continue to refer to Figure 3D , the second encoding submodule includes a second atomic feature extraction unit and a second molecular feature extraction unit, then the above S203-B2 includes: inputting the atomic characteristic matrix of the first product and the adjacency matrix under the i-th noise value into the first atomic feature extraction unit for atomic feature extraction, and obtaining an embedding (embedding) representation of each atom in the first product; then, inputting the embedding representation of each atom in the first product into the first molecular feature extraction unit for atomic feature interaction, and obtaining the first feature information of the first product under the i-th noise value, wherein the first feature information of the first product under the i-th noise value can be understood as the molecular structure characteristic information of the first product after being scrambled by the i-th noise value.

[0147] The embodiment of the present application does not limit the network structure of the p-th atomic characteristic extraction module, that is, the first atomic characteristic extraction module and the second atomic characteristic extraction module.

[0148] In one possible implementation, Figure 3E As shown in the figure, the p-th atomic feature extraction module includes M GNN layers, where M is a positive integer. The output value of the subsequent GNN layer is added to the predicted value of the previous GNN layer as the predicted value of the subsequent GNN layer, and the final output of the p-th atomic feature extraction module composed of M GNN layers is the sum of the predicted values ​​of each GNN layer.

[0149] At this time, the above S203-B02 inputs the adjacency matrix and node feature matrix of the target object into the p-th atom feature extraction unit, and obtains the embedded representation of each atom in the target object, including S203-B021 to S203-B023:

[0150] S203-B021. Input the adjacency matrix and node feature matrix of the target object into the p-th atom feature extraction unit to obtain the connection bond information of the j-th atom in the target object extracted by the m-th GNN layer, where m is a positive integer less than or equal to M.

[0151] The adjacency matrix of the embodiment of the present application includes four values ​​0, 1, 2, and 3, which represent no connection, single bond connection, double bond connection, and triple bond connection between atoms, respectively. The adjacency matrix and node feature matrix of the target object are input into the p-th atom feature extraction unit, so that the p-th atom feature extraction unit obtains the connection bond information of each atom in the target object extracted by the m-th GNN layer based on the adjacency matrix and node feature matrix of the target object. For ease of description, the following embodiment is explained by taking the connection bond information of the j-th atom in the target object as an example, and the process of determining the connection bond information of other atoms can refer to the process of determining the connection bond information of the j-th atom.

[0152] In one example, the adjacency matrix and node feature matrix of the target object are input into the pth atom feature extraction unit, and the information of the jth atom is aggregated with the information of neighboring atoms of different connection bond types of the jth atom through the mth GNN layer to obtain the connection bond information of the jth atom extracted by the mth GNN layer.

[0153] For example, Figure 3F As shown, taking the mth GNN layer as an example, assume that the jth atom has 4 neighbor atoms, of which the jth atom is connected to two neighbor atoms by single bonds, and the jth atom is connected to the other two neighbor atoms by double bonds. First, the adjacency matrix and node feature matrix of the target object are input into the mth GNN layer. For the jth atom in the target object, the mth GNN layer aggregates the information of the two neighbor atoms connected by single bonds of the jth atom with the information of the jth atom to obtain the single bond information of the jth atom extracted by the mth GNN layer. Similarly, the mth GNN layer aggregates the information of the two neighbor atoms connected by double bonds of the jth atom with the information of the jth atom to obtain the double bond information of the jth atom extracted by the mth GNN layer.

[0154] Among them, the process of extracting the connection bond information of the j-th atom in each GNN layer of M GNN layers is basically the same.

[0155] Optionally, the connection bond information of the jth atom in the target object extracted by the mth GNN layer can be determined according to the following formula (2):

[0156]

[0157] Among them, C represents the connection key type, A [C,·,·] represents the adjacency matrix corresponding to the connection key C, is the atomic feature matrix corresponding to the m-1th GNN layer, The connection bond information of the j-th atom in the target object extracted by the m-th GNN layer.

[0158] Formula (2) above means aggregating the information within each type of bond. For example, if an atom has four neighbors, two of which are double-bonded and the other two are single-bonded, the information about the double-bonded neighbors is aggregated first, followed by the single-bonded neighbors. In other words, the information is aggregated separately according to the bond type.

[0159] S203-B022. Fuse the connection bond information of the j-th atom with the m-1-th embedding representation of the j-th atom corresponding to the m-1-th GNN layer to obtain the m-th embedding representation of the j-th atom corresponding to the m-th GNN layer.

[0160] Continue to refer to Figure 3F As shown, the single bond information and double bond information of the j-th atom are combined with the m-1-th embedding representation of the j-th atom corresponding to the m-1-th GNN layer (i.e. Figure 3F After fusing and concatenating the mth embedding of the jth atom in the mth GNN layer, the MLP operation is performed to obtain the mth embedding representation of the jth atom corresponding to the mth GNN layer. Next, according to the above steps, the connection bond information of the jth atom extracted by the m+1th GNN layer is obtained, and the connection bond information of the jth atom extracted by the m+1th GNN layer is fused and concatenated with the mth embedding representation of the jth atom corresponding to the mth GNN layer. The MLP operation is then performed again to obtain the m+1th embedding representation of the jth atom corresponding to the m+1th GNN layer. This process is repeated until the Mth embedding representation of the jth atom corresponding to the Mth GNN layer is obtained.

[0161] In one example, the mth embedding representation of the jth atom corresponding to the mth GNN layer can be determined according to the following formula (3):

[0162]

[0163] in, is the m-1th embedding representation of the jth atom corresponding to the m-1th GNN layer, represents the connection bond information of the j-th atom extracted by the m-th GNN layer, is the mth embedding representation of the jth atom corresponding to the mth GNN layer.

[0164] S203-B023. Concatenate the embedding representations of the j-th atom corresponding to each of the M GNN layers to obtain the embedding representation of the j-th atom.

[0165] According to the above steps, the embedding representation of the j-th atom corresponding to each GNN layer in the M GNN layers can be obtained, and the embedding representation of the j-th atom corresponding to each GNN layer is concatenated to obtain the embedding representation of the j-th atom.

[0166] In one example, the embedded representation of the jth atom is determined according to the following formula (4):

[0167]

[0168] Among them, (Z out ) j is the embedding representation of the jth atom, is the embedding representation of the jth atom corresponding to the mth GNN layer.

[0169] In one example, the p-th atomic feature extraction unit can be simplified as the following formula (5):

[0170] Z out =MultiChannelGNN(A,Z in ) (5)

[0171] Among them, Z out represents the output of the p-th atomic feature extraction unit, A is the adjacency matrix of the input p-th atomic feature extraction unit, Z in is the node feature matrix input to the p-th atomic feature extraction unit. The above formula (5) indicates that the p-th atomic feature extraction unit is a multi-channel GNN, and the characteristic information extracted by each GNN layer is used as the output of the p-th atomic feature extraction unit.

[0172] The embodiment of the present application does not limit the network structure of the p-th molecular characteristic extraction module, that is, the first molecular characteristic extraction module and the second molecular characteristic extraction module.

[0173] After obtaining the embedded representation of each atom in the target according to the above method, execute S203-B02 to input the embedded representation of each atom in the target into the pth molecular feature extraction unit for feature interaction to obtain the first feature information of the target. For example, input the embedded representation of each atom in the first reactant into the first molecular feature extraction unit to obtain the first feature information x of the first reactant. R , input the embedded representation of each atom in the first product into the second molecular characteristic extraction unit to obtain the first characteristic information x of the first product under the i-th noise value p .

[0174] The above-mentioned first characteristic information can be understood as a low-dimensional vector representation of the target object.

[0175] The embodiment of the present application does not limit the network structure of the p-th molecular feature extraction unit.

[0176] In one example, the network structure of the p-th molecular feature extraction unit, ie, the first molecular characteristic extraction module and the second molecular characteristic extraction module, is a multi-layer perceptron MLP.

[0177] The above describes the model training process in conjunction with the network structure of the encoding module. The following describes the module training process in conjunction with the decoding module.

[0178] In some embodiments, as Figure 3G The decoding module includes a feature extraction unit and a feature decoding unit, wherein the feature extraction unit in the decoding module can be used to interactively learn the features extracted by the encoding module again, and the feature decoding unit can be understood as being used to decode the features extracted by the feature extraction unit to generate gradient field information of the adjacency matrix pointing to the first product.

[0179] exist Figure 3G Based on the network structure shown, the above S203-B2 includes S203-B21 to S203-B22:

[0180] S203-B21, input the first characteristic information of the first reactant and the first characteristic information of the first product at the i-th noise value into the feature extraction unit to obtain the second characteristic information of the first reactant and the second characteristic information of the first product at the i-th noise value respectively.

[0181] This application does not impose any restrictions on the network structure of the above-mentioned feature extraction unit.

[0182] In some embodiments, as Figure 3H As shown, the feature extraction unit includes a first feature extraction subunit and a second feature extraction subunit. In this case, the above S203-B21 includes:

[0183] S203-B211, inputting the first characteristic information of the first reactant into the first characteristic extraction subunit to obtain second characteristic information of the first reactant;

[0184] S203-B212: Input the first feature information of the first product under the i-th noise value into the second feature extraction subunit to obtain the second feature information of the first product under the i-th noise value.

[0185] This application does not limit the network structure of the first feature extraction subunit and the second feature extraction subunit.

[0186] S203-B22, inputting the second characteristic information of the first reactant and the second characteristic information of the first product under the i-th noise value into a feature decoding unit to obtain the gradient field information of the adjacency matrix of the first product under the i-th noise value.

[0187] In some embodiments, the first feature extraction subunit and the second feature extraction subunit are both transformer encoders (TransformerEncoder), and the feature decoding unit is a transformer decoder (TransformerDecoder). Specifically, the second characteristic information h of the first reactant extracted by the first feature extraction subunit is converted into R As query, the second feature information h of the first product extracted by the second feature extraction subunit under the i-th noise value is P As key and value, input them into TransformerDecoder together to obtain the gradient field information of the adjacency matrix of the first product under the i-th noise value.

[0188] In one example, the gradient field information S of the adjacency matrix of the first generated object under the i-th noise value can be obtained according to the following formula:

[0189] S=TransformerDecoder(h R ,h P ) (7)

[0190] in,

[0191] h R =TransformerEncoder(x R )

[0192] h P =TransformerEncoder(x P )

[0193] The above h R is the second characteristic information of the first reactant, h P is the second feature information of the first product under the i-th noise value.

[0194] Based on the network structure of the prediction model introduced in the above embodiment, in a specific embodiment, the prediction model and the training process of the prediction model involved in the embodiment of the present application are as follows: Figure 4 As shown, the adjacency matrix and atomic feature matrix of the first reactant are input into the first atomic feature extraction unit GNN encoder for atomic feature extraction, and the embedded representation of each atom in the first reactant is obtained, where the adjacency matrix of the first reactant ∈ R N×N×C , the atomic feature matrix of the first reactant ∈ RN×F , the GNN encoder includes M GNN layers. The adjacency matrix of the first product is subjected to noise processing, and the atomic feature matrix of the first product and the adjacency matrix after noise processing are input into the second atomic feature extraction unit GNN encoder for atomic feature extraction to obtain the embedded representation of each atom in the first product, where the atomic feature matrix of the first product ∈ R N×F , the adjacency matrix of the first product and the adjacency matrix after noise ∈ R N×N×C , the GNN encoder includes M GNN layers. Then, the embedded representation of each atom in the first reactant is input into the first molecular feature extraction unit MLP to extract the molecular features and obtain the first feature information x of the first reactant. R , where x R ∈R NxN The embedded representation of each atom in the first product is input into the second molecular feature extraction unit MLP to extract the molecular features, and the first feature information x of the first product under the i-th noise value is obtained. p , where x p ∈R NxN Next, the characteristic information x of the first reactant is R Input the first feature extraction subunit TransformerEncoder to obtain the second feature information h of the first reactant R , where h R ∈R NxN The first feature information x of the first product under the i-th noise value p Input the second feature extraction subunit to obtain the second feature information h of the first product under the i-th noise value p , where h p ∈R NxN Then, the second characteristic information h of the first reactant is R and the second characteristic information h of the first product under the i-th noise value p Input the feature decoding unit TransformerDecoder to obtain the gradient field information S∈R of the adjacency matrix of the first generated object predicted by the prediction model under the i-th noise value N×N .

[0195] The above description uses the prediction model to predict the gradient field information of the adjacency matrix of the first product under the i-th noise value. The prediction model predicts the gradient field information corresponding to the adjacency matrix of the first product under each noise value in the same manner. Specific reference is made to the description of the prediction model predicting the gradient field information of the adjacency matrix of the first product under the i-th noise value, which will not be repeated here.

[0196] According to the above steps, after the prediction model predicts the gradient field information of the adjacency matrix of the first product under different noise values, the following S204 is executed to implement the training of the prediction model.

[0197] S204 , training the prediction model according to the gradient field information of the adjacency matrix of the first product under different noise values ​​to obtain a trained prediction model.

[0198] The implementation methods of the above S204 include but are not limited to the following:

[0199] Method 1: Use the loss between the gradient field information of the adjacency matrix of the first product predicted by the prediction model under different noise values ​​and the true value of the gradient field information of the adjacency matrix of the first product under different noise values ​​to reversely train the prediction model to obtain a trained prediction model. Optionally, the true value of the gradient field information under different noise values ​​can be predicted by other trained gradient field information prediction models. For example, the first reactant and the first product under different noise values ​​are input into the gradient field information prediction model to obtain the gradient field information of the adjacency matrix of the first product under different noise values ​​output by the gradient field information prediction model, and the gradient field information of the adjacency matrix of the first product under different noise values ​​output by the gradient field information prediction model is used as the true value of the gradient field information of the adjacency matrix of the first product under different noise values.

[0200] Method 2: The above S204 includes S204-A1 and S204-A2:

[0201] S204-A1, determining the loss of the prediction model based on the gradient field information of the adjacency matrix of the first generated object under different noise values;

[0202] S204-A2: Adjust the parameters in the prediction model according to the loss of the prediction model to obtain a trained prediction model.

[0203] The embodiment of the present application does not limit the method of determining the loss of the prediction model.

[0204] In one example, the gradient field information of the adjacency matrix of the first product under different noise values ​​is substituted into the existing loss function to calculate the loss of the prediction model.

[0205] In one example, the loss of the prediction model is determined according to the following formula (8):

[0206]

[0207] Among them, θ represents the parameters of the prediction model, represents L noise values, The gradient field information of the adjacency matrix of the first product predicted by the prediction model under the i-th noise value, is the actual gradient field information of the adjacency matrix of the first product under the i-th noise value. The above loss function is used to constrain the gradient field information of the adjacency matrix of the first product predicted by the prediction model, so that the gradient field information predicted by the prediction model gradually approaches the true value of the gradient field information of the adjacency matrix of the first product.

[0208] In some embodiments, the actual gradient field information of the adjacency matrix of the first product under the i-th noise value can be inferred by adding noise. For example, if Gaussian distributed noise is used in the above S202, then At this point, the loss of the prediction model can be determined according to the following formula (9):

[0209]

[0210] As can be seen from the above, the present application adds different noise values ​​to the first product and uses the first product under different noise values ​​to jointly train the prediction model. The trained prediction model can predict the gradient field information of the adjacency matrix of the product (i.e., the transformation trend of the generation probability of the product), and then obtain the product based on the gradient field information of the adjacency matrix of the predicted product. For example, sampling is performed in the gradient field information of the adjacency matrix of the predicted product to obtain the adjacency matrix of the product, thereby realizing the prediction of the product.

[0211] In some embodiments, in order to facilitate the selection of prediction models under different noise values ​​when predicting the product subsequently, the embodiments of the present application also include: generating a correspondence between noise values ​​and model parameters, where the correspondence includes the model parameters of the prediction model corresponding to each noise value in the different noise values.

[0212] In some embodiments, if the p-th molecular feature extraction unit is an MLP layer, different model parameters are set in the MLP layer for different noise values. During model training, the model parameters corresponding to different noise values ​​in the MLP layer are reversely adjusted based on the calculated loss function to obtain the model parameters corresponding to different noise values ​​in the MLP layer. In this way, when subsequently using the prediction model corresponding to different noise values, the parameters in the MLP layer in the prediction model can be simply replaced with the model parameters corresponding to the target noise value, further improving the prediction efficiency of the model.

[0213] For example, the corresponding relationship between the noise value and the model parameters in the MLP layer is shown in Table 1:

[0214] <![CDATA[σ1]]> <![CDATA[α1]]> <![CDATA[β1]]> <![CDATA[σ2]]> <![CDATA[α2]]> <![CDATA[β2]]> ...... ...... ...... <![CDATA[σ L ]]> <![CDATA[α L ]]> <![CDATA[β L ]]>

[0215] Among them, α, β are the parameters of the MLP layer, such as the Lth noise value σ L The corresponding MLP layer parameter α L , β L Correspondingly, the other parameters in the prediction model are fixed and do not change with the noise value. In this way, when subsequently predicting the gradient field information of the adjacency matrix of the generated object under different noise values, the parameters in the MLP layer of the prediction model are determined as the parameters corresponding to the noise value, and the prediction model under the noise value can be obtained.

[0216] In some embodiments, when sampling is performed during the product prediction process, such as with noise annealing, it is desirable to quickly and easily obtain prediction models for different noise values. Thus, as shown in Table 1, if a prediction model trained at level σ1 is desired, inputting 1 will obtain the prediction model corresponding to σ1.

[0217] The model prediction method provided by the embodiment of the present application is by obtaining a first reactant and a first product of the first reactant; the first product is subjected to noise processing to obtain the first product under different noise values; the first reactant and the first product under different noise values ​​are input into the prediction model to obtain the gradient field information of the adjacency matrix of the first product output by the prediction model under different noise values; according to the gradient field information of the adjacency matrix of the first product under different noise values, the prediction model is jointly trained to obtain the trained prediction model. That is, the embodiment of the present application is subjected to joint training by the first reactant and the first product under different noise values ​​to the prediction model, and the prediction model under different noise values ​​can be obtained, the prediction model is used to predict the gradient field information of the adjacency matrix of the product of the second reactant, and the gradient field information is used to indicate the change trend of the generation probability of the product of the second reactant, so that when the product is predicted, the product can be accurately predicted according to the gradient field information of the adjacency matrix of the product. In addition, the prediction model of the present application does not limit the type of reactant, and thus can achieve the prediction of the products of different types of reactants.

[0218] The above describes the training process of the prediction model. The following describes the usage process of the prediction model.

[0219] Figure 5 A flow chart of a method for predicting a product provided in one embodiment of the present application is shown in FIG. Figure 5 Shown, including:

[0220] S501: Obtain a second reactant and K preset noise values.

[0221] The second reactant mentioned above can be understood as a reactant whose product needs to be predicted.

[0222] The above K is a positive integer less than or equal to L.

[0223] Optionally, the K noise values ​​are the same as some of the L noise values ​​used in the training of the prediction model.

[0224] Optionally, at least one of the K noise values ​​is different from the L noise values ​​used in training the prediction model.

[0225] S502 : For the i-th noise value among the K noise values, determine a prediction model corresponding to the i-th noise value, where the prediction model is trained using the above-mentioned model training method, wherein i is a positive integer from 1 to K.

[0226] As can be seen from the above embodiment, when training the prediction model, L noise values ​​are used to add noise to the training samples to obtain the prediction models under the L noise values. Based on this, in actual use, the prediction model under a specific noise value can be selected according to actual needs.

[0227] For example, for the i-th noise value among the K noise values, if the i-th noise value is the same as the noise value a among the L noise values ​​during the above training, the prediction model corresponding to the noise value a is determined as the prediction model corresponding to the i-th noise value.

[0228] For another example, if the i-th noise value is different from each of the L noise values ​​used in the above training, but is closest to the noise value b, the prediction model corresponding to the noise value b is determined as the prediction model corresponding to the i-th noise value.

[0229] It should be noted that during the above training, the prediction model is trained for each noise value to obtain a set of parameters. Therefore, the prediction model under a certain noise value described here can be understood as the parameters in the prediction model being the model parameters corresponding to the noise value.

[0230] In some embodiments, the above S502 includes steps S502-A1 to S502-A3:

[0231] S502-A1. Obtain the corresponding relationship between the noise value and the model parameter.

[0232] As shown in Table 1, the above correspondence includes the model parameters of the prediction model corresponding to each noise value in different noise values.

[0233] S502-A2: According to the i-th noise value, query the i-th group of model parameters corresponding to the i-th noise value from the corresponding relationship.

[0234] S502-A3: Use the i-th group of model parameters as parameters of the prediction model to obtain the prediction model corresponding to the i-th noise value.

[0235] The prediction model of this embodiment includes prediction models for different noise values. For example, when predicting the gradient field information for the i-th noise value, the prediction model for the i-th noise value is used. Specifically, the i-th set of model parameters corresponding to the i-th noise value is retrieved from the corresponding relationship shown in Table 1. The i-th set of model parameters is used as the parameters of the prediction model to obtain the prediction model corresponding to the i-th noise value. The prediction model corresponding to the i-th noise value is used to predict the gradient field information for the i-th noise value.

[0236] S503. Based on the target adjacency matrix of the second reactant and the second product at the i-1th noise value, sampling is performed in the gradient field information predicted by the prediction model corresponding to the i-th noise value to obtain the target adjacency matrix of the second product at the i-th noise value, where i is a positive integer from 1 to K, and the second product is the product of the second reactant.

[0237] The process of this step is a cyclic iterative process, for example, sorting the K noise values ​​from large to small. Starting from i=1, first randomly initialize an adjacency matrix for the second generator. As the target adjacency matrix of the second product under the 0th noise value. The initial adjacency matrix of the second product is equal to the size of the adjacency matrix of the second reactant, both of which are N×N matrices, where N is the number of atoms in the second reactant. The initial adjacency matrix is ​​a connection relationship of all atoms in the second product, so it can represent the structure of the second product. According to the second reactant and the initial adjacency matrix of the second product Sampling is performed from the gradient field information at the first noise value predicted by the prediction model to obtain the target adjacency matrix of the second product at the first noise value. Next, let i = i + 1, and based on the second reactant and the target adjacency matrix of the second product at the first noise value, sampling is performed from the gradient field information at the second noise value predicted by the prediction model to obtain the target adjacency matrix of the second product at the second noise value, and so on, until the target adjacency matrix of the second product at the Kth noise value is obtained. The Kth noise value is the minimum value among the above K noise values.

[0238] In the embodiment of the present application, the target adjacency matrix of the second product is sampled in the gradient field information predicted by the prediction model corresponding to K noise values. As the gradient field information changes, it gradually approaches the direction of the true adjacency matrix of the second product, and then the target adjacency matrix sampled from the gradient field information corresponding to the Kth noise value with the smallest noise value is used as the final adjacency matrix of the second product, thereby achieving accurate prediction of the product of the second reactant.

[0239] The method of obtaining the target adjacency matrix of the second product under the i-th noise value in the above S503 includes but is not limited to the following methods:

[0240] Method 1: Input the target adjacency matrix of the second reactant and the second product at the i-1th noise value into the prediction model corresponding to the i-th noise value, and obtain the gradient field information of the i-th target adjacency matrix of the second product output by the prediction model corresponding to the i-th noise value. The gradient field information reflects the change trend of the i-th target adjacency matrix of the second product. Therefore, sampling is performed in the gradient field information, for example, sampling is performed using annealing sampling or other sampling methods to obtain the target adjacency matrix of the second product at the i-th noise value.

[0241] In the second approach, the target adjacency matrix of the second product under the i-th noise value is generated based on T iterations. Specifically, the above S503 includes the following steps S503-A1 to S503-A3:

[0242] S503-A1. Input the t-1th adjacency matrix of the second product at the i-th noise value and the second reactant into the prediction model corresponding to the i-th noise value to obtain the t-1th gradient field information of the adjacency matrix of the second product at the i-th noise value, wherein t is a positive integer less than or equal to T. When t is 1, the t-1th adjacency matrix of the second product at the i-th noise value is the target adjacency matrix of the second product at the i-1th noise value.

[0243] S503-A2. Use the t-1 gradient field information of the adjacency matrix of the second generator at the i-th noise value to update the t-1-th adjacency matrix of the second generator at the i-th noise value to obtain the t-th adjacency matrix of the second generator at the i-th noise value, and repeat until t is T.

[0244] S503-A3: Determine the Tth adjacency matrix of the second product under the i-th noise value as the target adjacency matrix of the second product under the i-th noise value.

[0245] For example, the target adjacency matrix of the second product at the i-1th noise value and the second reactant are input into the prediction model corresponding to the i-th noise value to obtain the 0th gradient field information of the adjacency matrix of the second product at the i-1th noise value. The 0th gradient field information is used to update the target adjacency matrix of the second product at the i-1th noise value to obtain the 1st adjacency matrix of the second product at the i-th noise value. Next, the 1st adjacency matrix of the second product at the i-th noise value and the second reactant are input into the prediction model corresponding to the i-th noise value to obtain the 1st gradient field information of the adjacency matrix of the second product at the i-th noise value. The 1st gradient field information is used to update the 1st adjacency matrix of the second product at the i-th noise value to obtain the 2nd adjacency matrix of the second product at the i-th noise value. Next, the 2nd adjacency matrix of the second product at the i-th noise value and the second reactant are input into the prediction model corresponding to the i-th noise value to obtain the 2nd gradient field information of the adjacency matrix of the second product at the i-th noise value. The second adjacency matrix of the second product at the i-th noise value is updated using the second gradient field information to obtain the third adjacency matrix of the second product at the i-th noise value. Similarly, the T-1th adjacency matrix of the second product at the i-th noise value and the second reactant are input into the prediction model corresponding to the i-th noise value to obtain the T-1th gradient field information of the adjacency matrix of the second product at the i-th noise value. The T-1th gradient field information is used to update the T-1th adjacency matrix of the second product at the i-th noise value to obtain the T-th adjacency matrix of the second product at the i-th noise value. Finally, the T-th adjacency matrix of the second product at the i-th noise value is determined as the target adjacency matrix of the second product at the i-th noise value.

[0246] The above example takes the i-th noise value as an example. Referring to the above steps, the above steps are iteratively performed for each of the K noise values ​​until the target adjacency matrix of the second product under the K-th noise value is obtained.

[0247] In some embodiments, in the above S503-A2, the t-1th gradient field information of the adjacency matrix of the second product at the i-th noise value is used to update the t-1th adjacency matrix of the second product at the i-th noise value. Methods for obtaining the tth adjacency matrix of the second product at the i-th noise value include but are not limited to the following methods:

[0248] Method 1: The sum of the t-1 gradient field information of the adjacency matrix of the second product under the i-th noise value and the t-1-th adjacency matrix of the second product under the i-th noise value is used as the t-th adjacency matrix of the second product under the i-th noise value.

[0249] Method 2: The adjacency matrix update in S503-A2 includes the following steps:

[0250] Step A1: Determine the update step corresponding to the i-th noise value according to the K noise values ​​and the i-th noise value.

[0251] For example, the ratio of the i-th noise value to the K-th noise value is determined as the size of the update step corresponding to the i-th noise value.

[0252] For example, the update step size corresponding to the i-th noise value is determined according to the following formula (10):

[0253]

[0254] Among them, ∈ is the preset parameter, is the square of the Lth noise value, is the square of the i-th noise value, α i is the size of the update step corresponding to the i-th noise value.

[0255] Step A2: Determine the noise value corresponding to the t-th adjacency matrix.

[0256] The distribution of the noise value conforms to the normal distribution, for example, the noise value is z t , z t ~N(0,1), where “~” means obey or conform to, and N(0,1) represents the standard normal distribution.

[0257] Step A3: Determine the t-th adjacency matrix of the second product at the i-th noise value based on the t-1 gradient field information of the adjacency matrix of the second product at the i-th noise value, the t-1-th adjacency matrix of the second product at the i-th noise value, the update step corresponding to the i-th noise value, and the noise value corresponding to the t-th adjacency matrix.

[0258] In one example, the update step corresponding to the i-th noise value is multiplied by the noise value corresponding to the t-th adjacency matrix to obtain a first product. The t-1-th adjacency matrix of the second product at the i-th noise value, the t-1-th gradient field information of the adjacency matrix of the second product at the i-th noise value, and the first product are summed to obtain the t-th adjacency matrix of the second product at the i-th noise value.

[0259] In one example, the t-th adjacency matrix of the second reactant under the ith noise value can be determined according to the following formula (11):

[0260]

[0261] in, is the t-th adjacency matrix of the second generator under the ith noise value, is the t-1th adjacency matrix of the second product under the i-th noise value, R is the second reactant, is the adjacency matrix of the second generator, the t-1 gradient field information under the i-th noise value, α i is the noise value corresponding to the t-th adjacency matrix, z t is the size of the update step corresponding to the i-th noise value.

[0262] It should be noted that the above formula (11) is only an example. Equivalent deformation of the above formula (11), or adding, subtracting, multiplying or dividing one or more parameters from formula (11), also fall within the scope of protection of this application.

[0263] If the above i is 1, the target adjacency matrix of the second product under the i-1th noise value is the initial adjacency matrix of the second product Each element in conforms to the first normal distribution.

[0264] Since the connection keys in the above adjacency matrix are 0, 1, 2 or 3, in order to cover all connection chains, the variance of the first normal distribution is set to a positive number less than or equal to 3.

[0265] In some embodiments, the present application uses a Langevin annealing sampling method for sampling. The specific sampling process is as follows:

[0266] enter: (from large to small) ε, T, R, where ε is the minimum optimization step, T is the number of updates at each noise value, and R is the second reactant;

[0267] (1) Start (Randomly initialize an adjacency matrix);

[0268] (2) Noise level i←1 to K:

[0269] (3) (Set the update step size);

[0270] (4) Times t←1 to T: (update step number);

[0271] (5) Sampling z t ~N(0,1);

[0272] (6) (The sample follows The estimated gradient field information is updated by sliding);

[0273] (7) End the loop;

[0274] (8) (The obtained result is used as the initial adjacency matrix under the next noise level condition);

[0275] (9) End the loop;

[0276] (10) Return (Finally, the prediction of the generated adjacency matrix).

[0277] According to the Langevin annealing sampling method, sampling is performed on the gradient field information of the adjacency matrix of the second product predicted by the prediction model to obtain the target adjacency matrix of the second product under the Kth noise, and the second product is obtained according to the connection relationship of each atom in the target adjacency matrix.

[0278] In some embodiments, the initial adjacency matrix of the second product in step (1) above satisfies the following formula

[0279]

[0280] Among them, i and j are atoms in the second product, ε [i,j] ~N(0,3), that is, this application uses the part of the normal distribution N(0,3) greater than 0 to initialize each edge between atoms in the second product.

[0281] The above describes the specific sampling process involved in the embodiment of the present application. Now, in combination with the network structure of the prediction model, the process of inputting the t-1th adjacency matrix of the second product at the i-th noise value and the second reactant into the prediction model corresponding to the i-th noise value in the above S502-A41 to obtain the t-1th gradient field information of the second product at the i-th noise value is introduced.

[0282] In some embodiments, as Figure 6A As shown, the prediction model includes an encoding module and a decoding module. In this case, the above S503-A1 includes:

[0283] S503-A11, obtaining the adjacency matrix and node feature matrix of the second reactant;

[0284] S503-A12, such as Figure 6AAs shown, the adjacency matrix and node feature matrix of the second reactant (the node feature matrix is ​​also called the atomic feature matrix) and the t-1th adjacency matrix of the second product under the i-th noise value are input into the prediction model, so that the encoding module processes the adjacency matrix and node feature matrix of the second reactant and the t-1th adjacency matrix of the second product under the i-th noise value to obtain the first feature information of the second reactant and the first feature information of the t-1th adjacency matrix of the second product, so that the decoding module processes the first feature information of the second reactant and the first feature information of the t-1th adjacency matrix of the second product to obtain the t-1th gradient field information corresponding to the second product under the i-th noise value.

[0285] In some embodiments, as Figure 6B As shown, the encoding module includes a first encoding submodule and a second encoding submodule. The first encoding submodule is used to process the adjacency matrix and node feature matrix of the second reactant to obtain the first feature information of the second reactant. The second encoding submodule is used to process the t-1th adjacency matrix of the second product under the i-th noise value to obtain the first feature information of the t-1th adjacency matrix of the second product.

[0286] In some embodiments, as Figure 6C As shown, the first encoding submodule includes a first atomic feature extraction unit and a first molecular feature extraction unit.

[0287] The first atomic feature extraction unit is used to process the adjacency matrix and the node feature matrix of the second reactant to obtain an embedded representation of each atom in the second reactant.

[0288] The first molecular feature extraction unit is used to perform feature interaction on the embedded representation of each atom in the second reactant to obtain first feature information of the second reactant.

[0289] In some embodiments, referring to FIG6C , the second encoding submodule includes a second atomic feature extraction unit and a second molecular feature extraction unit.

[0290] The second atomic feature extraction unit is used to process the t-1th neighboring moment of the second product under the i-th noise value to obtain the embedded representation of each atom corresponding to the t-1th neighboring moment of the second product.

[0291] The first molecular feature extraction unit is used to perform feature interaction on the embedded representation of each atom corresponding to the t-1th adjacency matrix of the second product to obtain first feature information of the t-1th adjacency matrix of the second product.

[0292] In some embodiments, the network structure of the first atomic feature extraction unit and the second atomic feature extraction unit includes M GNN layers as shown in 3E, and each GNN layer is based on Figure 3F The method shown determines the embedded representation of atoms. The specific process is described in the above training process embodiment and will not be repeated here.

[0293] In some embodiments, as shown in FIG Figure 6D As shown, the decoding module includes a feature extraction unit and a feature decoding unit;

[0294] The feature extraction unit is configured to obtain the second characteristic information of the second reactant and the second characteristic information of the t-1th adjacency matrix of the second product based on the first characteristic information of the second reactant and the first characteristic information of the t-1th adjacency matrix of the second product;

[0295] The feature decoding unit is used to obtain the t-1th gradient field information corresponding to the second product under the i-th noise value according to the second characteristic information of the second reactant and the second feature information of the t-1th adjacency matrix of the second product.

[0296] In some embodiments, as Figure 6E As shown, the feature extraction unit includes a first feature extraction subunit and a second feature extraction subunit;

[0297] The first feature extraction subunit is used to obtain second characteristic information of the second reactant based on the first characteristic information of the second reactant;

[0298] The second feature extraction subunit is used to obtain the second feature information of the t-1th adjacency matrix of the second product according to the first feature information of the t-1th adjacency matrix of the second product.

[0299] In a possible implementation, the first feature extraction subunit and the second feature extraction subunit are both encoders in the converter, and the feature decoding unit is a decoder in the converter.

[0300] Optionally, the first molecular feature extraction unit and the second molecular feature extraction unit are MLP layers. In this case, the corresponding relationship between the noise value and the model parameter is the corresponding relationship between the noise value and the parameters in the MLP layer.

[0301] It should be noted that the model used to predict the t-1th gradient field information corresponding to the second product under the i-th noise value is the prediction model trained by the above-mentioned training method, wherein the specific network structure of the prediction model refers to the specific introduction in the above-mentioned training steps and will not be repeated here.

[0302] In a specific embodiment, the prediction model and the prediction process of the prediction model involved in the embodiment of the present application are as follows: Figure 6FAs shown, Figure 6F As shown, the above S502-A412 includes: inputting the adjacency matrix and atomic feature matrix of the second reactant into the first atomic feature extraction unit GNN encoder to extract atomic features, and obtaining an embedded representation of each atom in the second reactant, wherein the adjacency matrix of the second reactant ∈ R N×N×C , the atomic feature matrix of the second reactant ∈ R N×F , the GNN encoder includes M GNN layers. The t-1th adjacency matrix of the second product under the i-th noise value is input into the second atomic feature extraction unit GNN encoder for atomic feature extraction, and the embedded representation of each atom corresponding to the t-1th adjacency matrix of the second product is obtained. Then, the embedded representation of each atom in the second reactant is input into the molecular feature extraction unit MLP to obtain the first feature information x of the second reactant. R , where x R ∈R NxN The embedded representation of each atom corresponding to the t-1th adjacency matrix of the second product is input into the second molecular feature extraction unit MLP to extract the molecular features, and the first feature information x of the t-1th adjacency matrix of the second product under the i-th noise value is obtained. p , where x p ∈R NxN Next, the characteristic information x of the second reactant is R Input the first feature extraction subunit TransformerEncoder to obtain the second feature information h of the second reactant R , where h R ∈R NxN The first characteristic information x of the t-1th adjacency matrix of the second product under the i-th noise value is p Input the second feature extraction subunit to obtain the second feature information h of the t-1th adjacency matrix of the second product under the i-th noise value p , where h p ∈R NxN Then, the second characteristic information h of the second reactant is R The second characteristic information h of the t-1th adjacency matrix of the second product under the i-th noise value p Input the feature decoding unit TransformerDecoder to obtain the t-1th gradient field information S∈R corresponding to the second product predicted by the prediction model under the i-th noise value N×N .

[0303] According to the above method, after obtaining the t-1th gradient field information corresponding to the second product predicted by the prediction model at the i-th noise value, the t-1th gradient field information corresponding to the second product at the i-th noise value is used to update the t-1th adjacency matrix of the second product at the i-th noise value to obtain the t-th adjacency matrix of the second product at the i-th noise value, and the process is repeated until t is T. The T-th adjacency matrix of the second product at the i-th noise value is determined as the target adjacency matrix of the second product at the i-th noise value.

[0304] The above steps are performed for each of the K noise values, iteratively, until the target adjacency matrix of the second product under the K-th noise value is determined. The K-th noise value is the minimum noise value among the K noise values. Then, the following S503 is executed.

[0305] S504 : Predicting a second product according to the target adjacency matrix of the second product under the K-th noise value, wherein the K-th noise value is the minimum value among the K noise values.

[0306] For example, based on the connection relationship between the atoms in the target adjacency matrix of the second product under the Kth noise value, the product corresponding to the target adjacency matrix under the Lth noise value can be obtained, and the product is determined as the product of the second reactant, that is, the second product.

[0307] The product prediction method provided in the embodiments of the present application uses a prediction model to predict the gradient field information of the adjacency matrix of the second product corresponding to the second reactant. This gradient field information represents the changing trend of the generation probability of the second product. The adjacency matrix of the second product is then sampled within the gradient field information of the adjacency matrix of the second product to obtain the final adjacency matrix of the second product, thereby accurately predicting the second product corresponding to the second reactant. The second reactant can be any type of reactant, that is, the embodiments of the present application can predict products of any type of reactant.

[0308] The preferred embodiments of the present application are described in detail above in conjunction with the accompanying drawings. However, the present application is not limited to the specific details in the above embodiments. Within the technical concept of the present application, a variety of simple modifications can be made to the technical solution of the present application, and these simple modifications all fall within the scope of protection of the present application. For example, the various specific technical features described in the above specific embodiments can be combined in any suitable manner unless there is any contradiction. In order to avoid unnecessary repetition, the present application will not further explain various possible combinations. For another example, the various different embodiments of the present application can also be arbitrarily combined, and as long as they do not violate the ideas of the present application, they should also be regarded as the contents disclosed in the present application.

[0309] It should also be understood that in the various method embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0310] Combined with the above Figure 2 6, the method embodiment of the present application is described in detail, and the following is combined with Figures 7 to 9 , describe in detail the device embodiments of the present application.

[0311] Figure 7 A schematic diagram of a structure of a training device for a prediction model provided in an embodiment of the present application. The training device 20 can be a computing device or a component of a computing device (e.g., an integrated circuit, a chip, etc.) for executing the above-mentioned model training method.

[0312] An acquisition unit 21 is configured to acquire a first reactant and a first product of the first reactant;

[0313] a noise adding unit 22, configured to perform noise processing on the first product to obtain first products with different noise values;

[0314] A prediction unit 23 is configured to input the first reactant and the first product under different noise values ​​into a prediction model to obtain gradient field information of an adjacency matrix of the first product under different noise values ​​output by the prediction model;

[0315] The training unit 24 is used to train the prediction model based on the gradient field information of the adjacency matrix of the first product under different noise values ​​to obtain the trained prediction model, wherein the trained prediction model is used to predict the gradient field information of the adjacency matrix of the product of the second reactant, and the gradient field information is used to indicate the changing trend of the generation probability of the product of the second reactant.

[0316] In some embodiments, the above-mentioned training unit 24 is specifically used to determine the loss of the prediction model based on the gradient field information of the adjacency matrix of the first generated object under different noise values; according to the loss of the prediction model, adjust the parameters in the prediction model to obtain the trained prediction model.

[0317] In some embodiments, the noise adding unit 22 is specifically configured to obtain an adjacency matrix of the first product; perform noise processing on the adjacency matrix of the first product to obtain adjacency matrices of the first product under different noise values;

[0318] Correspondingly, the prediction unit 23 is specifically used to obtain the adjacency matrix and node feature matrix of the first reactant, as well as the node feature matrix of the first product; the adjacency matrix and node feature matrix of the first reactant, the node feature matrix of the first product and the adjacency matrix under different noise values ​​are input into the prediction model to obtain the gradient field information of the adjacency matrix of the first product under different noise values ​​output by the prediction model.

[0319] In some embodiments, the prediction model includes an encoding module and a decoding module, and the prediction unit 23 is specifically configured to input the adjacency matrix and the node feature matrix of the first reactant, and the atomic property matrix and the adjacency matrix of the first product at the i-th noise value into the encoding module for the i-th noise value among the different noise values, to obtain the first feature information of the first reactant and the first feature information of the first product at the i-th noise value output by the encoding module;

[0320] The first characteristic information of the first reactant and the first characteristic information of the first product under the i-th noise value are input into the decoding module to obtain the gradient field information of the first generated adjacency matrix under the i-th noise value output by the decoding module.

[0321] In some embodiments, the encoding module includes a first encoding submodule and a second encoding submodule, and a prediction unit 23, which is specifically used to input the adjacency matrix and node feature matrix of the first reactant into the first encoding submodule to obtain the first feature information of the first reactant; input the atomic property matrix and the adjacency matrix of the first product under the i-th noise value into the second encoding submodule to obtain the first feature information of the first product under the i-th noise value.

[0322] In some embodiments, the p-th encoding submodule includes a p-th atomic feature extraction unit and a p-th molecular feature extraction unit, and the prediction unit 23 is further used to input the adjacency matrix and the node feature matrix of the target object into the p-th atomic feature extraction unit to obtain an embedded representation of each atom in the target object; input the embedded representation of each atom in the target object into the p-th molecular feature extraction unit for feature interaction to obtain first feature information of the target object;

[0323] Among them, when p is 1, the pth encoding submodule is the first encoding submodule, and the adjacency matrix and node feature matrix of the target object are the adjacency matrix and node feature matrix of the first reactant; when p is 2, the pth encoding submodule is the second encoding submodule, and the adjacency matrix and node feature matrix of the target object are the atomic property matrix of the first product and the adjacency matrix under the i-th noise value.

[0324] In some embodiments, the p-th atom feature extraction module includes M graph neural network GNN layers, where M is a positive integer, and the prediction unit 23 is specifically used to input the adjacency matrix and node feature matrix of the target object into the p-th atom feature extraction unit to obtain the connection bond information of the j-th atom in the target object extracted by the m-th GNN layer, where m is a positive integer less than or equal to M; the connection bond information of the j-th atom and the m-1-th embedding representation of the j-th atom corresponding to the m-1-th GNN layer are fused to obtain the m-th embedding representation of the j-th atom corresponding to the m-1-th GNN layer; the embedding representation of the j-th atom corresponding to each GNN layer in the M GNN layers is spliced ​​to obtain the embedding representation of the j-th atom.

[0325] In some embodiments, the prediction unit 23 is specifically used to input the adjacency matrix and node feature matrix of the target object into the p-th atom feature extraction unit, and aggregate the information of the j-th atom with the information of neighboring atoms of different connection bond types through the m-th GNN layer to obtain the connection bond information of the j-th atom about different connection bond types extracted by the m-th GNN layer.

[0326] In some embodiments, the decoding module includes a feature extraction unit and a feature decoding unit, and a prediction unit 23, which is specifically used to input the first feature information of the first reactant and the first feature information of the first product at the i-th noise value into the feature extraction unit to obtain the second feature information of the first reactant and the second feature information of the first product at the i-th noise value; input the second feature information of the first reactant and the second feature information of the first product at the i-th noise value into the feature decoding unit to obtain the gradient field information of the adjacency matrix of the first product at the i-th noise value.

[0327] In some embodiments, the feature extraction unit includes a first feature extraction subunit and a second feature extraction subunit, and a prediction unit 23, which is specifically used to input the first feature information of the first reactant into the first feature extraction subunit to obtain the second characteristic information of the first reactant; and input the first feature information of the first product under the i-th noise value into the second feature extraction subunit to obtain the second feature information of the first product under the i-th noise value.

[0328] In some embodiments, the first feature extraction subunit and the second feature extraction subunit are both encoders in the converter, and the feature decoding unit is a decoder in the converter.

[0329] In some embodiments, the noise adding unit 22 is specifically configured to add different noise values ​​to the adjacency matrix of the first product by adding noise using Gaussian distribution, so as to obtain the adjacency matrix of the first product under different noise values.

[0330] In some embodiments, the training unit 24 is further configured to generate a correspondence between noise values ​​and model parameters, wherein the correspondence includes model parameters of the prediction model corresponding to each noise value among different noise values.

[0331] Optionally, the p-th molecular feature extraction unit is MLP.

[0332] Optionally, the correspondence between the noise value and the model parameter is the correspondence between the noise value and the parameter in the MLP.

[0333] It should be understood that the device embodiment and the method embodiment may correspond to each other, and similar descriptions may refer to the method embodiment. To avoid repetition, they will not be described here. Specifically, Figure 7 The training device shown can correspond to the corresponding subject in executing the model training method of the embodiment of the present application, and the aforementioned and other operations and / or functions of each module in the training device are respectively for realizing the corresponding processes in each method in the above-mentioned model training. For the sake of brevity, they will not be repeated here.

[0334] Figure 8 A schematic diagram of a structure of a product prediction device provided in an embodiment of the present application. The prediction device 30 can be a computing device or a component of a computing device (e.g., an integrated circuit, a chip, etc.), which is used to execute the above-mentioned product prediction method.

[0335] An acquisition unit 31 is configured to obtain a second reactant and K preset noise values, where K is a positive integer less than or equal to L;

[0336] a determining unit 32 configured to determine, for an i-th noise value among the K noise values, a prediction model corresponding to the i-th noise value, wherein the prediction model is obtained by training using the above-mentioned training method, and i is a positive integer from 1 to K;

[0337] a sampling unit 33 for sampling, based on the target adjacency matrix of the second reactant and the second product at the (i-1)th noise value, the gradient field information predicted by the prediction model corresponding to the (i)th noise value, to obtain the target adjacency matrix of the second product at the (i)th noise value, where the second product is the product of the second reactant;

[0338] The prediction unit 34 is configured to determine the second product according to a target adjacency matrix of the second product under a K-th noise value, wherein the K-th noise value is a minimum value among the K noise values.

[0339] In some embodiments, the determining unit 32 is specifically configured to obtain a correspondence between noise values ​​and model parameters, the correspondence including model parameters of the prediction model corresponding to each noise value among different noise values; query, based on the i-th noise value, from the correspondence, an i-th group of model parameters corresponding to the i-th noise value; and use the i-th group of model parameters as parameters of the prediction model to obtain a prediction model corresponding to the i-th noise value;

[0340] The sampling unit 33 is specifically used to sample the gradient field information predicted by the prediction model corresponding to the i-th noise value according to the second reactant and the target adjacency matrix of the second reactant at the i-1-th noise value, so as to obtain the target adjacency matrix of the second reactant at the i-th noise value.

[0341] In some embodiments, the sampling unit 33 is specifically configured to input the t-1th adjacency matrix of the second product at the i-th noise value and the second reactant into the prediction model corresponding to the i-th noise value to obtain the t-1th gradient field information of the adjacency matrix of the second product at the i-th noise value, wherein t is a positive integer less than or equal to T, and when t is 1, the t-1th adjacency matrix of the second product at the i-th noise value is the target adjacency matrix of the second product at the i-1 noise value; use the t-1th gradient field information of the adjacency matrix of the second product at the i-th noise value to update the t-1th adjacency matrix of the second product at the i-th noise value to obtain the t-th adjacency matrix of the second product at the i-th noise value, and repeat the process until t is T; and determine the T-th adjacency matrix of the second product at the i-th noise value as the target adjacency matrix of the second reactant at the i-th noise value.

[0342] In some embodiments, the sampling unit 33 is specifically configured to determine the size of the update step corresponding to the i-th noise value based on the K noise values ​​and the i-th noise value; determine the noise value corresponding to the t-th adjacency matrix; and determine the t-th adjacency matrix of the second product under the i-th noise value based on the t-1 gradient field information corresponding to the adjacency matrix of the second product under the i-th noise value, the t-1 adjacency matrix of the second product under the i-th noise value, the size of the update step corresponding to the i-th noise value, and the noise value corresponding to the t-th adjacency matrix.

[0343] In some embodiments, the prediction model includes an encoding module and a decoding module, and the above-mentioned sampling unit 33 is specifically used to obtain the adjacency matrix and node feature matrix of the second reactant; the adjacency matrix and node feature matrix of the second reactant, and the t-1th adjacency matrix of the second product under the i-th noise value are input into the prediction model, so that the encoding module processes the adjacency matrix and node feature matrix of the second reactant, and the t-1th adjacency matrix of the second product under the i-th noise value to obtain the first feature information of the second reactant and the first feature information of the t-1th adjacency matrix of the second product, so that the decoding module processes the first feature information of the second reactant and the first feature information of the t-1th adjacency matrix of the second product to obtain the t-1th gradient field information of the adjacency matrix of the second product under the i-th noise value.

[0344] In some embodiments, the encoding module includes a first encoding submodule and a second encoding submodule;

[0345] The first encoding submodule is used to process the adjacency matrix and the node feature matrix of the second reactant to obtain first feature information of the second reactant;

[0346] The second encoding submodule is used to process the t-1th adjacency matrix of the second product under the i-th noise value to obtain first feature information of the t-1th adjacency matrix of the second product.

[0347] In some embodiments, the first encoding submodule includes a first atomic feature extraction unit and a first molecular feature extraction unit;

[0348] The first atomic feature extraction unit is used to process the adjacency matrix and the node feature matrix of the second reactant to obtain an embedded representation of each atom in the second reactant;

[0349] The first molecular feature extraction unit is used to perform feature interaction on the embedded representation of each atom in the second reactant to obtain first feature information of the second reactant.

[0350] In some embodiments, the second encoding submodule includes a second atomic feature extraction unit and a second molecular feature extraction unit;

[0351] The second atom feature extraction unit is used to process the t-1th adjacency matrix of the second product under the i-th noise value to obtain an embedded representation of each atom corresponding to the t-1th adjacency matrix of the second product;

[0352] The second molecular feature extraction unit is used to perform feature interaction on the embedded representation of each atom corresponding to the t-1th adjacency matrix of the second product to obtain first feature information of the t-1th adjacency matrix of the second product.

[0353] In some embodiments, the decoding module includes a feature extraction unit and a feature decoding unit;

[0354] The feature extraction unit is used to obtain second characteristic information of the second reactant and second feature information of the t-1th adjacency matrix of the second product according to the first feature information of the second reactant and the first feature information of the t-1th adjacency matrix of the second product;

[0355] The feature decoding unit is used to obtain the t-1th gradient field information of the adjacency matrix of the second product under the i-th noise value based on the second characteristic information of the second reactant and the second feature information of the t-1th adjacency matrix of the second product.

[0356] In some embodiments, the feature extraction unit includes a first feature extraction subunit and a second feature extraction subunit;

[0357] The first feature extraction subunit is used to obtain second characteristic information of the second reactant based on the first feature information of the second reactant;

[0358] The second feature extraction subunit is configured to obtain second feature information of the t-1th adjacency matrix of the second product based on the first feature information of the t-1th adjacency matrix of the second product.

[0359] In some embodiments, the first feature extraction subunit and the second feature extraction subunit are both encoders in the converter, and the feature decoding unit is a decoder in the converter.

[0360] Optionally, if i is 1, each element in the target adjacency matrix of the second product under the i-1th noise value conforms to the first normal distribution.

[0361] Optionally, the variance of the first normal distribution is a positive number less than or equal to 3.

[0362] It should be understood that the device embodiment and the method embodiment may correspond to each other, and similar descriptions may refer to the method embodiment. To avoid repetition, they will not be described here. Specifically, Figure 8 The prediction device shown can correspond to the corresponding subject in executing the prediction method in the embodiment of the present application, and the aforementioned and other operations and / or functions of each module in the prediction device are respectively for implementing the corresponding processes in the above-mentioned prediction method. For the sake of brevity, they will not be repeated here.

[0363] The apparatus of the embodiment of the present application is described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that the functional module can be implemented in hardware form, can be implemented by instructions in software form, and can also be implemented by a combination of hardware and software modules. Specifically, the steps of the method embodiment in the embodiment of the present application can be completed by the integrated logic circuit of the hardware in the processor and / or the instructions in software form. The steps of the method disclosed in conjunction with the embodiment of the present application can be directly embodied as being executed by the hardware processor, or can be executed by a combination of hardware and software modules in the processor. Optionally, the software module can be located in a mature storage medium in the art such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps in the above method embodiment in conjunction with its hardware.

[0364] Figure 9 This is a block diagram of a computing device involved in an embodiment of the present application, which is used to execute the method described in the above embodiment. For details, please refer to the description in the above method embodiment.

[0365] Figure 9 The computing device 200 shown includes a memory 201, a processor 202, and a communication interface 203. The memory 201, processor 202, and communication interface 203 are communicatively connected to each other. For example, the memory 201, processor 202, and communication interface 203 may be communicatively connected using a network connection. Alternatively, the computing device 200 may further include a bus 204. The memory 201, processor 202, and communication interface 203 are communicatively connected to each other via the bus 204. Figure 9 The computing device 200 is a computer device 200 in which a memory 201 , a processor 202 , and a communication interface 203 are connected to each other via a bus 204 .

[0366] The memory 201 may be a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 201 may store programs. When the program stored in the memory 201 is executed by the processor 202, the processor 202 and the communication interface 203 are used to perform the above method.

[0367] The processor 202 may be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), a graphics processing unit (GPU), or one or more integrated circuits.

[0368] The processor 202 may also be an integrated circuit chip with signal processing capabilities. During implementation, the method of the present application may be implemented by hardware integrated logic circuits in the processor 202 or by software instructions. The processor 202 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor may be a microprocessor or any conventional processor. The software module may be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, or the like. The storage medium is located in the memory 201, and the processor 202 reads the information in the memory 201 and, in combination with its hardware, implements the method of the embodiment of the present application.

[0369] The communication interface 203 uses a transceiver module such as, but not limited to, a transceiver to implement communication between the computing device 200 and other devices or communication networks. For example, a data set can be obtained through the communication interface 203.

[0370] When the computing device 200 includes a bus 204 , the bus 204 may include a path for transmitting information between various components of the computing device 200 (eg, the memory 201 , the processor 202 , and the communication interface 203 ).

[0371] The present application also provides a computer storage medium having a computer program stored thereon, which, when executed by a computer, enables the computer to perform the method of the above-mentioned method embodiment. In other words, the present application also provides a computer program product containing instructions, which, when executed by a computer, enables the computer to perform the method of the above-mentioned method embodiment.

[0372] The present application also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method of the above method embodiment.

[0373] In other words, when implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a digital video disc (DVD)), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0374] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0375] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is merely a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0376] Modules described as separate components may or may not be physically separate, and components displayed as modules may or may not be physical modules, i.e., they may be located in one place or distributed across multiple network elements. Some or all of the modules may be selected based on actual needs to achieve the purpose of the present embodiment. For example, the functional modules in the various embodiments of the present application may be integrated into a processing module, or each module may exist physically separately, or two or more modules may be integrated into a single module.

[0377] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here. In addition, reference can be made between the various method embodiments and the various device embodiments, and the same or corresponding contents in different embodiments can be referenced to each other without further elaboration.

Claims

1. A model training method, characterized in that: include: obtaining a first reactant and a first product of the first reactant; performing noise processing on the first product to obtain first products with different noise values; Inputting the first reactant and the first product under different noise values ​​into a prediction model, and obtaining gradient field information of the adjacency matrix of the first product under different noise values ​​output by the prediction model; The loss of the prediction model is determined based on the gradient field information of the adjacency matrix of the first product under different noise values, and the parameters in the prediction model are adjusted based on the loss of the prediction model to obtain the trained prediction model, wherein the trained prediction model is used to predict the gradient field information of the adjacency matrix of the product of the second reactant, and the gradient field information is used to indicate the changing trend of the generation probability of the product of the second reactant.

2. The method according to claim 1, characterized in that The step of adding noise to the first product to obtain first products with different noise values ​​includes: Get the adjacency matrix of the first product; performing noise processing on the adjacency matrix of the first product to obtain adjacency matrices of the first product under different noise values; The step of inputting the first reactant and the first product under different noise values ​​into a prediction model to obtain gradient field information of the adjacency matrix of the first product output by the prediction model corresponding to different noise values ​​includes: Obtaining an adjacency matrix and a node feature matrix of the first reactant, and a node feature matrix of the first product; The adjacency matrix and node feature matrix of the first reactant, the node feature matrix and the adjacency matrix under different noise values ​​of the first product are input into the prediction model to obtain the gradient field information of the adjacency matrix of the first product under different noise values ​​output by the prediction model.

3. The method according to claim 2, characterized in that The prediction model includes an encoding module and a decoding module. The adjacency matrix and node feature matrix of the first reactant, the node feature matrix and the adjacency matrix under different noise values ​​of the first product are input into the prediction model to obtain gradient field information of the adjacency matrix of the first product under different noise values ​​output by the prediction model, including: For an i-th noise value among the different noise values, inputting the adjacency matrix and the node feature matrix of the first reactant, and the atomic property matrix and the adjacency matrix of the first product at the i-th noise value into the encoding module, and obtaining first feature information of the first reactant and first feature information of the first product at the i-th noise value output by the encoding module; The first characteristic information of the first reactant and the first characteristic information of the first product under the i-th noise value are input into the decoding module to obtain the gradient field information of the first generated adjacency matrix under the i-th noise value output by the decoding module.

4. The method according to claim 3, characterized in that The encoding module includes a first encoding submodule and a second encoding submodule, and inputting the adjacency matrix and the node feature matrix of the first reactant, and the atomic property matrix and the adjacency matrix of the first product under the i-th noise value into the encoding module, and obtaining the first feature information of the first reactant and the first feature information of the first product under the i-th noise value output by the encoding module, including: Inputting the adjacency matrix and the node feature matrix of the first reactant into the first encoding submodule to obtain first feature information of the first reactant; The atomic property matrix of the first product and the adjacency matrix under the i-th noise value are input into the second encoding submodule to obtain first characteristic information of the first product under the i-th noise value.

5. The method according to claim 4, characterized in that The p-th encoding submodule includes a p-th atomic feature extraction unit and a p-th molecular feature extraction unit, and the method further includes: Inputting the adjacency matrix and node feature matrix of the target object into the p-th atom feature extraction unit to obtain an embedded representation of each atom in the target object; Inputting the embedded representation of each atom in the target object into the p-th molecular feature extraction unit for feature interaction to obtain first feature information of the target object; Among them, when p is 1, the pth encoding submodule is the first encoding submodule, and the adjacency matrix and node feature matrix of the target object are the adjacency matrix and node feature matrix of the first reactant; when p is 2, the pth encoding submodule is the second encoding submodule, and the adjacency matrix and node feature matrix of the target object are the atomic property matrix of the first product and the adjacency matrix under the i-th noise value.

6. The method according to claim 5, characterized in that The p-th atom feature extraction module includes M graph neural network (GNN) layers, where M is a positive integer. The adjacency matrix and node feature matrix of the target object are input into the p-th atom feature extraction unit to obtain an embedded representation of each atom in the target object, including: Inputting the adjacency matrix and node feature matrix of the target object into the p-th atom feature extraction unit to obtain the connection bond information of the j-th atom in the target object extracted by the m-th GNN layer, where m is a positive integer less than or equal to M; Fusing the connection bond information of the j-th atom and the m-1-th embedding representation of the j-th atom corresponding to the m-1-th GNN layer to obtain the m-th embedding representation of the j-th atom corresponding to the m-th GNN layer; The embedded representations of the j-th atom corresponding to each GNN layer in the M GNN layers are concatenated to obtain the embedded representation of the j-th atom.

7. The method according to claim 6, characterized in that The step of inputting the adjacency matrix and the node feature matrix of the target object into the p-th atom feature extraction unit to obtain the connection bond information of the j-th atom in the target object extracted by the m-th GNN layer includes: The adjacency matrix and node feature matrix of the target object are input into the p-th atom feature extraction unit, and the information of the j-th atom is aggregated with the information of neighboring atoms of different connection bond types through the m-th GNN layer to obtain the connection bond information of the j-th atom about different connection bond types extracted by the m-th GNN layer.

8. The method according to claim 3, characterized in that The decoding module includes a feature extraction unit and a feature decoding unit. The first feature information of the first reactant and the first feature information of the first product under the i-th noise value are input into the decoding module, and the gradient field information of the adjacency matrix of the first product under the i-th noise value is obtained by the decoding module. The method includes: Inputting the first characteristic information of the first reactant and the first characteristic information of the first product at the i-th noise value into the feature extraction unit to obtain the second characteristic information of the first reactant and the second characteristic information of the first product at the i-th noise value; The second characteristic information of the first reactant and the second characteristic information of the first product under the i-th noise value are input into the feature decoding unit to obtain the gradient field information of the adjacency matrix of the first product under the i-th noise value.

9. The method according to claim 8, characterized in that The feature extraction unit includes a first feature extraction subunit and a second feature extraction subunit, and inputting the first feature information of the first reactant and the first feature information of the first product at the i-th noise value into the feature extraction unit to obtain the second feature information of the first reactant and the second feature information of the first product at the i-th noise value includes: inputting the first characteristic information of the first reactant into the first characteristic extraction subunit to obtain second characteristic information of the first reactant; The first feature information of the first product at the i-th noise value is input into the second feature extraction subunit to obtain the second feature information of the first product at the i-th noise value.

10. The method according to any one of claims 1 to 9, characterized in that The method further comprises: A correspondence between noise values ​​and model parameters is generated, wherein the correspondence includes the model parameters of the prediction model corresponding to each noise value among the different noise values.

11. A method for predicting a product, characterized in that: include: Obtaining a second reactant to be predicted and K preset noise values, where K is a positive integer less than or equal to L; For an i-th noise value among the K noise values, determining a prediction model corresponding to the i-th noise value, wherein the prediction model is obtained by training using the training method according to any one of claims 1 to 10, where i is a positive integer from 1 to K; According to the target adjacency matrix of the second reactant and the second product at the (i-1)th noise value, sampling is performed in the gradient field information predicted by the prediction model corresponding to the (i)th noise value to obtain the target adjacency matrix of the second product at the (i)th noise value, where the second product is the product of the second reactant; The second product is determined according to a target adjacency matrix of the second product under a K-th noise value, wherein the K-th noise value is a minimum value among the K noise values.

12. The method according to claim 11, characterized in that Determining the prediction model corresponding to the i-th noise value includes: Obtaining a correspondence between noise values ​​and model parameters, wherein the correspondence includes model parameters of the prediction model corresponding to each noise value among different noise values; According to the i-th noise value, querying the i-th group of model parameters corresponding to the i-th noise value from the corresponding relationship; The i-th group of model parameters is used as parameters of the prediction model to obtain the prediction model corresponding to the i-th noise value.

13. The method according to claim 12, characterized in that The method of sampling the gradient field information predicted by the prediction model corresponding to the i-th noise value based on the target adjacency matrix of the second reactant and the second product at the i-1th noise value to obtain the target adjacency matrix of the second product at the i-th noise value includes: Inputting the t-1th adjacency matrix of the second product at the i-th noise value and the second reactant into the prediction model corresponding to the i-th noise value to obtain the t-1th gradient field information of the adjacency matrix of the second product at the i-th noise value, wherein t is a positive integer less than or equal to T, and when t is 1, the t-1th adjacency matrix of the second product at the i-th noise value is the target adjacency matrix of the second product at the i-1th noise value; Using the t-1 gradient field information of the adjacency matrix of the second product at the i-th noise value, updating the t-1 th adjacency matrix of the second product at the i-th noise value to obtain the t th adjacency matrix of the second product at the i-th noise value, and repeating the process until t equals T; The T th adjacency matrix of the second product under the i th noise value is determined as the target adjacency matrix of the second reactant under the i th noise value.

14. The method according to claim 13, characterized in that The method uses the t-1 gradient field information of the adjacency matrix of the second product under the ith noise value to update the t-1th adjacency matrix of the second product under the ith noise value to obtain the tth adjacency matrix of the second product under the ith noise value, including: Determining the size of the update step corresponding to the i-th noise value according to the K noise values ​​and the size of the i-th noise value; Determine the noise value corresponding to the t-th adjacency matrix; The t-th adjacency matrix of the second generated object at the i-th noise value is determined based on the t-1 gradient field information corresponding to the adjacency matrix of the second generated object at the i-th noise value, the t-1-th adjacency matrix of the second generated object at the i-th noise value, the size of the update step corresponding to the i-th noise value, and the noise value corresponding to the t-th adjacency matrix.

15. The method according to claim 13, characterized in that The prediction model includes an encoding module and a decoding module, and the step of inputting the t-1th adjacency matrix of the second product under the ith noise value and the second reactant into the prediction model corresponding to the ith noise value to obtain the t-1th gradient field information of the adjacency matrix of the second product under the ith noise value includes: Obtain the adjacency matrix and node feature matrix of the second reactant; The adjacency matrix and node feature matrix of the second reactant, and the t-1th adjacency matrix of the second product under the i-th noise value are input into the prediction model, so that the encoding module processes the adjacency matrix and node feature matrix of the second reactant, and the t-1th adjacency matrix of the second product under the i-th noise value to obtain the first feature information of the second reactant and the first feature information of the t-1th adjacency matrix of the second product, so that the decoding module processes the first feature information of the second reactant and the first feature information of the t-1th adjacency matrix of the second product to obtain the t-1th gradient field information of the adjacency matrix of the second product under the i-th noise value.

16. A model training device, characterized in that: include: an acquisition unit, configured to acquire a first reactant and a first product of the first reactant; a noise adding unit, configured to perform noise adding processing on the first product to obtain first products with different noise values; A prediction unit, configured to input the first reactant and the first product under different noise values ​​into a prediction model, and obtain gradient field information of an adjacency matrix of the first product under different noise values ​​output by the prediction model; A training unit is used to determine the loss of the prediction model based on the gradient field information of the adjacency matrix of the first product under different noise values, and adjust the parameters in the prediction model according to the loss of the prediction model to obtain the trained prediction model, wherein the trained prediction model is used to predict the gradient field information of the adjacency matrix of the product of the second reactant, and the gradient field information is used to indicate the change trend of the adjacency matrix of the product of the second reactant.

17. A prediction device for a product, characterized in that: include: an acquisition unit, configured to obtain a second reactant and K preset noise values; a determining unit, configured to determine, for an i-th noise value among the K noise values, a prediction model corresponding to the i-th noise value, wherein the prediction model is obtained by training using the training method according to any one of claims 1 to 10, where i is a positive integer from 1 to K; a sampling unit, configured to sample, based on the target adjacency matrix of the second reactant and the second product at the (i-1)th noise value, the gradient field information predicted by the prediction model corresponding to the (i)th noise value, to obtain the target adjacency matrix of the second product at the (i)th noise value, where the second product is a product of the second reactant; A prediction unit is configured to determine the second product according to a target adjacency matrix of the second product under a K-th noise value, wherein the K-th noise value is a minimum value among the K noise values.

18. A computing device, characterized in that include: processor and memory; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the method according to any one of claims 1 to 10 or 11 to 15.

19. A computer-readable storage medium, characterized in that The storage medium includes computer instructions, and when the instructions are executed by a computer, the computer is caused to implement the method according to any one of claims 1 to 10 or 11 to 15.

Images