Chemical reaction prediction system and its control method, and learning method for chemical reaction prediction system

Abstract

The present invention relates to a chemical reaction prediction system, a control method thereof, and a learning method for a chemical reaction prediction system. More specifically, the present invention relates to a chemical reaction prediction system that performs forward reaction prediction based on electron flow, a control method thereof, and a learning method for a chemical reaction prediction system.

Description

Technical Field

[0001] The present invention relates to a chemical reaction prediction system, a control method thereof, and a learning method of the chemical reaction prediction system. More specifically, the present invention relates to a chemical reaction prediction system that performs forward reaction prediction based on an electron flow, a control method thereof, and a learning method of the chemical reaction prediction system.

Background Art

[0002] With the development of artificial intelligence, the number of cases where excellent results are achieved using artificial intelligence technology in various fields has increased rapidly.

[0003] In particular, recently, in the field of natural science, efforts to solve various scientific problems by utilizing artificial intelligence technology have been continuously carried out. For example, in the field of chemistry, research such as predicting the results of chemical reactions between molecules or designing new molecules using artificial intelligence technology has been actively conducted.

[0004] In this regard, organic synthesis is one of the important issues in new drug development and materials science, and predicting the results of chemical reactions is very important in the design of new molecules.

[0005] Conventionally, in order to predict this, a sequence-based model using SMILES (Simplified Molecular Input Line Entry System) strings has been utilized. However, since the conversion by SMILES strings cannot naturally represent the molecular structure in a graph form, there is a limit in that the graph structure of the molecule cannot be fully reflected.

[0006] To address these problems, various graph-based approaches have been proposed that represent molecules as graphs. Graph representations accurately depict molecular structures and provide intuitive information for interpreting and predicting chemical reaction mechanisms. In particular, research utilizing graph neural networks (GNNs) is actively being conducted, which has the advantage of accurately modeling the interactions between atoms within a molecule. [Overview of the project] [Problems that the invention aims to solve]

[0007] The present invention provides a chemical reaction prediction system that performs forward reaction prediction based on electron flow, a control method therefor, and a learning method for the chemical reaction prediction system.

[0008] Furthermore, the present invention provides a chemical reaction prediction system that performs interpretable forward reaction prediction using an electron-flow-inspired graph diffusion model, a control method thereof, and a learning method for the chemical reaction prediction system.

[0009] The present invention provides a chemical reaction prediction system and a control method thereof, as well as a learning method for the chemical reaction prediction system, which can deeply understand molecular structures and accurately predict chemical reactions between molecular structures based on that understanding.

[0010] Furthermore, the present invention provides a chemical reaction prediction system and a control method thereof that can understand chemical reaction mechanisms and predict various types of chemical reaction results, as well as a learning method for the chemical reaction prediction system.

[0011] More specifically, the present invention provides a chemical reaction prediction model based on electron flow using graph diffusion, in which both the input and output structures are formed in graph space.

[0012] Furthermore, the present invention provides a method for training a chemical reaction prediction model that enables a deeper understanding of chemical reaction mechanisms and the generation of more accurate and interpretable chemical reaction prediction results. [Means for solving the problem]

[0013] A chemical reaction prediction method performed cooperatively by a memory and processor according to the present invention may include the steps of: receiving information on a plurality of molecular structures as input to an encoder; obtaining an embedding vector corresponding to the plurality of molecular structures using the information on the plurality of molecular structures in the embedding layer of the encoder; performing an attention calculation related to the interaction between atoms of the plurality of molecular structures in a multi-head self-attention layer and updating the embedding vector based on the calculation; storing the updated embedding vector in the update step in memory and inputting the updated embedding vector stored in memory to a decoder; performing bond prediction and atom prediction using the updated embedding vector in the decoder, which are predicted as chemical reactions of the plurality of molecular structures; and obtaining a final chemical reaction result predicted from the chemical reactions of the plurality of molecular structures using the results of the bond prediction and atom prediction.

[0014] In an embodiment, the step of obtaining the chemical reaction result comprises the steps of: sampling the initialization reaction result using the bond prediction result and the atom prediction result; stabilizing the sampled initialization reaction result through a diffusion feedback process; and obtaining the final chemical reaction result stabilized through the diffusion feedback process.

[0015] In an embodiment, the method further includes the step of obtaining a molecular graph relating to the plurality of molecular structures by converting atoms into nodes and interatomic bonds into edges, wherein the plurality of molecular structures include a first molecular structure and a second molecular structure, and the step of obtaining the molecular graph further includes the step of obtaining a first molecular graph including nodes and edges corresponding to the first molecular structure by converting atoms constituting the first molecular structure into nodes and interatomic bonding relationships into edges using a pre-specified graph transformation algorithm, and the step of obtaining a second molecular graph including nodes and edges corresponding to the second molecular structure by converting atoms constituting the second molecular structure into nodes and interatomic bonding relationships into edges using the pre-specified graph transformation algorithm.

[0016] In an embodiment, the information relating to the plurality of molecular structures includes information relating to nodes and edges corresponding to the first molecular structure and information relating to nodes and edges corresponding to the second molecular structure, and the embedding vector is characterized in that it includes at least one piece of information relating to the atomic type, atomic charge, number of hydrogens, number of radical electrons, and degree of the nodes corresponding to the first molecular structure and the nodes corresponding to the second molecular structure.

[0017] In one embodiment, the step of updating the embedding vector is characterized by adding different biases to the attention score calculated in the Multi-Head Self-Attention layer, depending on the type of connection between the nodes constituting the first molecular graph and the second molecular graph.

[0018] The mechanism is characterized in that the bond type includes single bond type, double bond type, triple bond type, and aromatic bond type.

[0019] In an embodiment, the method further includes the step of extracting at least one of the following: an adjacency matrix, a bond type matrix, a shortest path matrix, and K-hop neighbors, corresponding to the first molecular graph and the second molecular graph, using nodes and edges corresponding to the first molecular graph and nodes and edges corresponding to the second molecular graph, wherein the adjacency matrix includes information on direct connections between nodes constituting the first molecular graph and the second molecular graph; the bond type matrix includes information on bond types between nodes constituting the first molecular graph and the second molecular graph; the shortest path matrix includes information on the shortest path lengths between nodes constituting the first molecular graph and the second molecular graph; and the K-hop neighbors include information on neighboring nodes within K hops for each node constituting the first molecular graph and the second molecular graph.

[0020] In one embodiment, the step of updating the embedding vector is characterized by inputting the output vector of the multi-head self-attention layer to a feed-forward neural network layer, updating the output vector of the multi-head self-attention layer using at least one of the adjacency matrix, bond type matrix, shortest path matrix, and K-hop neighbor information in the feed-forward neural network layer, and identifying the output vector of the feed-forward neural network layer as the updated embedding vector.

[0021] In this embodiment, the bond prediction is performed by executing a dot product using the updated embedding vector, and the dot product is performed for each pair of atoms corresponding to the updated embedding vector.

[0022] In the embodiment, the step of performing bond prediction includes obtaining dot product values ​​for each of the plurality of bond types for each of the atom pairs based on the dot product calculation, wherein the plurality of bond types are associated with at least one of single bond, double bond, bond formation, bond collapse, and no change.

[0023] In this embodiment, the step of performing bond prediction further includes the steps of generating a probability distribution for each of the plurality of bond types for each of the atom pairs using the dot product value obtained by the dot product calculation, and obtaining a transformation matrix that predicts the change in the bond state of each of the atom pairs using the probability distribution.

[0024] In an embodiment, in the step of generating the probability distribution, for each atom pair, by applying a softmax function to the inner product values obtained for the plurality of bond types, the probability distribution of the plurality of bond types for each atom pair is generated.

[0025] In an embodiment, in the step of performing the atom prediction, the method further includes: using a softmax output layer to generate an atomic property probability distribution of each atom corresponding to the updated embedding vector; and using the probability distribution to predict the atomic property of the atom corresponding to the updated embedding vector, where the atomic property includes the charge state of the atom that can change in the chemical reaction process of the plurality of molecular structures.

[0026] In an embodiment, the diffusion feedback process repeatedly evaluates each bond transformation of the initial chemical reaction product and removes or changes unstable bonds.

[0027] In an embodiment, the diffusion feedback process uses a transformation probability matrix and a target transformation matrix to evaluate the predicted bond transformation for each step of the repeated execution in order to predict the change in the bond state of the initial chemical reaction product, and generates the final chemical reaction product using an interpolation factor.

[0028] The chemical reaction prediction method executed in cooperation by a memory and a processor according to the present invention includes: receiving information regarding a plurality of molecular structures as an input to an encoder (Encoder); obtaining a molecular graph using atoms as nodes and bonds as edges based on the plurality of molecular structures; obtaining an embedding vector corresponding to the molecular graph using the information regarding the plurality of molecular structures in an embedding layer of the encoder; performing an attention operation related to the interaction between atoms of the plurality of molecular structures in a multi-head self-attention layer, and updating the embedding vector based on the operation; storing the embedding vector updated by the update step in a memory, and inputting the updated embedding vector stored in the memory to a decoder; performing bond prediction and atom prediction, which are predicted as chemical reactions of the plurality of molecular structures, using the updated embedding vector in the decoder; obtaining a final chemical reaction result predicted from the chemical reactions of the plurality of molecular structures using the results of the bond prediction and the results of the atom prediction; calculating a loss function between the final chemical reaction result and label data including the actual bond state and atom state corresponding to the plurality of molecular structures; and optimizing parameters of at least one of the encoder and the decoder to minimize the loss function.

[0029] A chemical reaction prediction system according to the present invention, comprising a memory, an encoder, a decoder, and at least one processor, wherein the encoder receives information on a plurality of molecular structures, obtains a molecular graph based on the plurality of molecular structures using atoms as nodes and bonds as edges, obtains an embedding vector corresponding to the molecular graph in the embedding layer of the encoder, performs an attention calculation related to the interaction between atoms of the plurality of molecular structures in the multi-head self-attention layer of the encoder, updates the embedding vector based on the calculation, the processor stores the updated embedding vector in the update step in memory, inputs the updated embedding vector stored in memory to the decoder, the decoder uses the updated embedding vector to perform bond prediction and atom prediction, which are predicted as chemical reactions of the plurality of molecular structures, and obtains the final chemical reaction result predicted from the chemical reactions of the plurality of molecular structures using the results of the bond prediction and atom prediction.

[0030] A program, which is executed by one or more processes in an electronic device according to the present invention and stored on a computer-readable medium, may include instructions to cause an encoder to perform the following steps: receiving information about a plurality of molecular structures as input; obtaining a molecular graph based on the plurality of molecular structures using atoms as nodes and bonds as edges; obtaining an embedding vector corresponding to the molecular graph in the embedding layer of the encoder; performing an attention calculation related to the interaction between atoms of the plurality of molecular structures in a multi-head self-attention layer and updating the embedding vector based on the calculation; storing the updated embedding vector in memory and inputting the updated embedding vector stored in memory to a decoder; performing bond prediction and atom prediction, which are predicted as chemical reactions of the plurality of molecular structures, using the updated embedding vector in the decoder; and obtaining a final chemical reaction result predicted from the chemical reactions of the plurality of molecular structures using the results of the bond prediction and atom prediction. [Effects of the Invention]

[0031] As described above, the chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system according to the present invention allow for a deep understanding of the chemical reaction mechanism and accurate prediction of the chemical reaction results by modeling electron transfer through the graph structure of molecules.

[0032] Furthermore, the chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system according to the present invention provide a chemical reaction prediction model that can deeply understand the chemical reaction mechanism and accurately predict the chemical reaction results, thereby reducing the time and cost of experiments for users. This reduces research and development costs and shortens the time to market for new products.

[0033] On the other hand, according to the chemical reaction prediction system, its control method, and learning method for the chemical reaction prediction system of the present invention, by performing both the input of molecular structures and the output for predicted results in graph space, forward reaction prediction can be performed regardless of the permutation and order of SMILES.

[0034] Furthermore, according to the chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system of the present invention, it is possible to ensure consistency between transformations by enabling simultaneous sampling of multiple interdependent transformations that occur in parallel within the molecular graph.

[0035] Furthermore, according to the chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system of the present invention, in order to solve problems that may arise due to symmetrical structures, it is possible to break the symmetry by including noise or a sampling mechanism and form an effective output structure, thereby preventing the occurrence of ineffective configurations. [Brief explanation of the drawing]

[0036] [Figure 1] This is a conceptual diagram illustrating a response generation system to which the chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system according to the present invention are applied. [Figure 2a] This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 2b] This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 2c]This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 2d] This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 2e] This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 2f] This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 3] This is a flowchart illustrating the learning method for the chemical reaction prediction model according to the present invention. [Figure 4a] This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 4b] This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 4c] This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 4d] This is a conceptual diagram illustrating the chemical reaction prediction model according to the present invention. [Figure 5a] This is a conceptual diagram illustrating an example of the application of the chemical reaction prediction model according to the present invention in a response system. [Figure 5b] This is a conceptual diagram illustrating an example of the application of the chemical reaction prediction model according to the present invention in a response system. [Figure 5c] This is a conceptual diagram illustrating an example of the application of the chemical reaction prediction model according to the present invention in a response system. [Figure 5d] This is a conceptual diagram illustrating an example of the application of the chemical reaction prediction model according to the present invention in a response system. [Figure 5e] This is a conceptual diagram illustrating an example of the application of the chemical reaction prediction model according to the present invention in a response system. [Figure 6] This is a conceptual diagram illustrating an example of the application of the chemical reaction prediction model according to the present invention in a response system. [Figure 7]This is a conceptual diagram illustrating an example of the application of the chemical reaction prediction model according to the present invention in a response system. [Modes for carrying out the invention]

[0037] The embodiments disclosed herein will be described in detail below with reference to the accompanying drawings, but regardless of the reference numerals used in the drawings, identical or similar components will be given the same reference numerals, and redundant descriptions thereof will be omitted. The suffixes “module” and “part” used for components in the following description are added or mixed for the sake of ease of writing the specification and do not have any distinguishing meaning or role in themselves. Furthermore, when describing the embodiments disclosed herein, if it is determined that a detailed description of the relevant prior art would obscure the gist of the embodiments disclosed herein, such detailed description will be omitted. In addition, the accompanying drawings are intended solely to facilitate the understanding of the embodiments disclosed herein, and the technical ideas disclosed herein should not be limited by the accompanying drawings and should be understood to include all modifications, equivalents and substitutions that fall within the concept and technical scope of the present invention.

[0038] Terms including ordinal numbers such as "1st," "2nd," etc., may be used to describe various components, but the components are not limited to those defined by these terms. These terms are used solely to distinguish one component from another.

[0039] When it is stated that one component is “connected” or “linked” to another component, it should be understood that it may be directly connected or linked to the other component, but there may also be another component between them. On the other hand, when it is stated that one component is “directly connected” or “directly linked” to another component, it should be understood that there is no other component between them.

[0040] A singular expression includes plural forms unless otherwise clearly indicated in the context.

[0041] In this application, terms such as “includes” or “having” are intended to specify the presence of features, figures, steps, actions, components, parts, or combinations thereof as described in the specification, and should be understood not to preemptively exclude the possibility of the presence or addition of one or more other features, figures, steps, actions, components, parts, or combinations thereof.

[0042] The present invention will be discussed in more detail below with reference to the attached drawings. Figure 1 is a conceptual diagram illustrating a response generation system to which the chemical reaction prediction system and its control method, as well as the learning method for the chemical reaction prediction system, according to the present invention, are applied. Figures 2a, 2b, 2c, 2d, 2e, and 2f are conceptual diagrams illustrating the chemical reaction prediction model according to the present invention, and Figure 3 is a flowchart illustrating the learning method for the chemical reaction prediction model according to the present invention. Furthermore, Figures 4a, 4b, 4c, and 4d are conceptual diagrams illustrating the chemical reaction prediction model according to the present invention, and Figures 5a, 5b, 5c, 5d, 5e, 6, and 7 are conceptual diagrams illustrating examples of applications in a response system to which the chemical reaction prediction model according to the present invention is applied.

[0043] The chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system according to the present invention can be usefully utilized in a variety of situations. Predicting chemical reactions has various applications, for example, it can be usefully utilized in research for designing new substances or developing new drugs. In this regard, organic synthesis is one of the important challenges in new drug development and materials science, and predicting the results of chemical reactions is extremely important for designing new molecules. The present invention predicts chemical reactions by graphing molecular structures, and provides a system, control method, and learning method for predicting chemical reactions between molecular structures. The chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system according to the present invention are embodied based on a "chemical reaction prediction model," and in the following, for the sake of convenience of explanation, the system, method, and learning method will not be named separately but will be collectively referred to as the "chemical reaction prediction model."

[0044] On the other hand, the chemical reaction prediction model according to the present invention can be applied to various industries and services, and can be usefully utilized, for example, in answer generation systems based on language models. In recent years, with the development of deep learning technology, generative artificial intelligence (AI) technology has been attracting attention. More specifically, generative AI models can generate new data in various forms, such as text, images, and audio, from given data, which offers a level of applicability that differs from mere classification or prediction of existing data. The chemical reaction prediction model according to the present invention can also be applied to such answer generation systems and can be usefully utilized in various fields where chemical reaction prediction is necessary, such as research for designing new substances or developing new drugs.

[0045] A response generation system to which the chemical reaction prediction model according to the present invention can be applied will be briefly examined with reference to Figure 1. The response generation system shown in Figure 1 may include various prediction and analysis models, and may be a system that generates prediction results for physical properties of molecular structures or a system that designs molecules having properties desired by the user. Furthermore, the response generation system according to the present invention may be a system that generates prediction results for chemical reactions of new types of molecules and / or between multiple molecules. In addition, the response generation system according to the present invention may be a system that generates prediction results for the deformation of existing materials and the synthesis of various materials (e.g., new materials, polymer materials, nanomaterials, composite materials, organic materials, pharmaceutical materials, etc.).

[0046] The answer generation system according to the present invention includes a super-large-scale foundational model (or super-large-scale foundational artificial intelligence model, or generative artificial intelligence model), and the present invention aims to improve the efficiency of natural science research by minimizing the risk of failure in research. Such an answer generation system is also called an answer generation platform based on a super-large-scale foundational model. However, "super-large-scale foundational model" is also called a generative model, foundational model, or Large Language Model (LLM).

[0047] On the other hand, as shown in Figure 1, the response generation system 100 may include at least one of the following: an input unit 110, an output unit 120, a communication unit 130, a storage unit 140, and a super-large-scale foundational model 200. Here, the super-large-scale foundational model 200 is also called a foundational model, and the foundational model may mean a super-large-scale AI core foundational model trained on a massive dataset.

[0048] Although not shown in the figures, the answer generation system 100 may include one or more processors, such processors may include one or more general-purpose processors and / or one or more special-purpose processors (e.g., digital signal processors, tensor processing units (TPUs), graphics processing units (GPUs), neural network processing units (NPUs), application-specific integrated circuits, application-specific semiconductors (ASICs), etc.). One or more processors may be configured to execute instruction words, computer-readable instructions, and / or other instructions described herein that are stored (or included) in the memory unit 140. Such an answer generation system and method can perform data processing described later in cooperation with the memory and at least one processor. The processor can perform a series of operations and data processing using the data and information stored in the memory. In this case, the memory may be configured as the memory unit 140.

[0049] On the other hand, the input unit 110 is a means of data input and may be configured in various ways. For example, the input unit 110 may be configured to receive user input. The input unit 110 may be configured to receive user input from the user terminal 10. Here, "receiving input" may mean receiving an input signal (or selection signal) corresponding to the user input, based on the user input being made through the configuration of the input unit provided in the user terminal 10.

[0050] The input unit 110 is also called a user interface module. The input unit 110 may include a touchscreen, computer mouse, keyboard, keypad, touchpad, trackball, joystick, voice recognition module, or other similar device. However, the present invention does not limit the type of input unit 110. Furthermore, in the present invention, the input unit 110 does not necessarily mean hardware means, but can be understood as a channel for receiving input from the user.

[0051] Here, user input may include documents, text, images (or videos), audio, etc. In this case, the response generation system 100 may further include a module that converts audio to text.

[0052] Next, the output unit 120 can output information through the configuration of an output unit (e.g., display unit, touchscreen, speaker, etc.) provided on the user terminal 10, which is linked to the answer generation system 100. For example, the output unit 120 can output a page 1000 (or service page) linked to the answer generation system 100 to the display unit of the user terminal 10. Furthermore, the output unit 120 does not necessarily refer to hardware means, but can be understood as a channel for outputting results to the user.

[0053] The communication unit 130 may then be connected wirelessly or via a wired network to the user terminal 10, servers (e.g., a central server, an external server, etc.), devices, and at least one network, and may be configured to receive or transmit overall data and information necessary for the operation of the response generation system 100.

[0054] Here, the user terminal 10 may include at least one of the following: mobile phone, smartphone, notebook computer, laptop computer, slate PC, tablet PC, ultrabook, desktop computer, digital broadcasting terminal, PDA (personal digital assistant), PMP (portable multimedia player), navigation system, and wearable device (e.g., smartwatch, smart glass, HMD (head mounted display)).

[0055] Furthermore, the communication unit 130 can support various communication methods depending on the communication standard of the device it communicates with.

[0056] For example, the communication unit 130 supports WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Wi-Fi (Wireless Fidelity) Direct, DLNA (Digital Living Network Alliance), WiBro (Wireless Broadband), WiMAX (World Interoperability for Microwave Access), HSDPA (High Speed ​​Downlink Packet Access), HSUPA (High Speed ​​Uplink Packet Access), LTE (Long Term Evolution), LTE-A (Long Term Evolution-Advanced), 5G (5th Generation Mobile Telecommunication), and Bluetooth (registered trademark). TM It may be configured to communicate with a target using at least one of the following technologies: RFID (Radio Frequency Identification), Infrared Data Association (IrDA), UWB (Ultra-Wideband), ZigBee, NFC (Near Field Communication), Wi-Fi Direct, or Wireless USB (Wireless Universal Serial Bus).

[0057] On the other hand, the storage unit 140 plays a role in storing various data related to the present invention and may include one or more non-temporary computer-readable storage media that can be read and / or accessed by at least one of one or more processors.

[0058] One or more computer-readable storage media may include volatile and / or non-volatile storage components such as optical, magnetic, organic, or other memory or disk storage devices. In some examples, the storage unit 140 can be embodied using a single physical device (e.g., one optical, magnetic, organic, or other memory or disk storage device), while in other examples, the storage unit 140 can be embodied using two or more physical devices.

[0059] The storage unit 140 may include computer-readable instructions and additional data. The storage unit 140 may include storage necessary to perform at least some of the methods, scenarios and techniques described herein and / or at least some of the functions of the apparatus and network.

[0060] Furthermore, at least a portion of the storage unit 140 may be cloud storage or a cloud server. The storage unit 140 may store at least a portion of the data corresponding to user input received from the input unit 110 and the training data.

[0061] In other words, the memory unit 140 only needs to be a space where the information necessary for the operation of the answer generation system 100 is stored, and it can be understood that there are no physical space constraints.

[0062] On the other hand, the ultra-large-scale base model 200 may be configured to generate prediction results for physical properties of molecular structures or to design molecules with properties desired by the user. Furthermore, the ultra-large-scale base model 200 can generate prediction results for new types of molecules and / or chemical reactions between multiple molecules, or generate prediction results for the deformation of existing materials and the synthesis of various materials (e.g., new materials, polymer materials, nanomaterials, composite materials, organic materials, pharmaceutical materials, etc.).

[0063] In this regard, the ultra-large-scale foundational model 200 may include at least one of the following: a document understanding model 300, a chemical reaction prediction model 400, and a molecular property prediction model 500.

[0064] The Document Understanding model 300 can extract various forms of content from documents (e.g., papers, books, patent documents, reports, etc.) that meet predefined content criteria. More specifically, the Document Understanding model 300 may be a model trained to understand structured data, unstructured data, linguistic data (or linguistic elements), and non-linguistic data (or non-linguistic elements) contained in a document, and to extract various content (or data) and knowledge based on its understanding.

[0065] Here, the pre-set content criteria can be set in various ways and may be determined according to the intended use and purpose of the response generation system 100 according to the present invention.

[0066] For example, if the purpose of using the response generation system 100 is chemistry, biotechnology, new materials, new substances, and new drug development, the document understanding model 300 may be trained to understand and extract content related to chemistry, biotechnology, new materials, new substances, and new drug development from the documents being analyzed.

[0067] In this case, the pre-defined content criteria may include content related to molecular structure concerning at least one of the following: chemistry, biotechnology, new materials, new substances, and new drug development. Here, the document comprehension model 300 can extract content related to chemistry, biotechnology, new materials, new substances, and new drug development from the document under analysis based on the pre-defined content criteria.

[0068] For the sake of clarity, this specification describes the predefined content criteria as relating to at least one of the following: chemistry, biotechnology, new materials, new substances, and new drug development, but is not necessarily limited to these.

[0069] The document comprehension model 300 can extract at least one of the following from the document under analysis that satisfies the pre-defined content criteria: text, molecular structure, mathematical formula, chart, table, and image.

[0070] For example, the document understanding model 300 can understand the chemical structure of the molecular structure contained in the document 20 to be analyzed, and based on the understanding, it can convert the molecular structure into SMILES string notation and extract it. Furthermore, the document understanding model 300 can understand the chemical structure of the molecular structure and, based on the understanding, perform a graph transformation corresponding to the molecular structure.

[0071] On the other hand, in this invention, inputting "molecular structure" can be understood as inputting information that can identify a molecule. In this case, the information that can identify a molecular structure may take various forms, such as molecular structural formulas, molecular graphs, chemical formulas, molecular structural formulas according to SMILES notation, and molecular structural images.

[0072] Furthermore, the document understanding model 300 can understand the text 23 related to molecular structure 21 from among the texts contained in the document 20 to be analyzed, and extract this as text data 23 related to molecular structure 21.

[0073] Furthermore, the document understanding model 300 can recognize the rows and columns that make up a table 24 related to the molecular structure 21 from the document 20 to be analyzed, and convert and extract this into structured data 25 in a format such as HTML or Excel.

[0074] Furthermore, the document understanding model 300 can extract relationship information (or relationships) between molecular structures contained in the document.

[0075] As described above, the document understanding model 300 can convert and extract various forms of data contained in a document into data in a form that the ultra-large-scale infrastructure model 200 can understand (for example, machine-readable data). The data extracted using the document understanding model 300 may be sorted by page or document and stored in the storage unit 140 (or memory). In the present invention, the document understanding model 300 is also called a deep document understanding model.

[0076] Next, the chemical reaction prediction model 400 may be configured to predict the chemical reaction between the input molecular structures (or compounds or reactants) and output (or generate) the predicted chemical reaction result (or product). For example, as shown in Figure 2a, when information about the molecular structures 401a and 401b of a particular compound is input, the chemical reaction prediction model 400 can predict the product 402 that will be formed by the chemical reaction between the molecular structures 401a and 401b of that particular compound.

[0077] The chemical reaction prediction model 400 according to the present invention can model electron transfer through the graph structure of molecules. The chemical reaction prediction model 400 according to the present invention may also be a prediction model based on electron flow using graph diffusion that can deeply understand the chemical reaction mechanism and accurately predict the chemical reaction results. The chemical reaction prediction model 400 can represent both input and output in graph space. By learning the graph transformation of chemical reactions, the chemical reaction prediction model 400 can output predicted results of reaction products and / or by-products based on the graph structure of the input compound. In the present invention, the chemical reaction prediction model 400 is also called an electron flow model, or an electron flow-based forward reaction prediction model, or an electron-flow inspired graph diffusion model for interpretable forward reaction prediction.

[0078] On the other hand, the chemical reaction prediction model 400 described above may include at least one of the encoder 410 and the decoder 420.

[0079] The encoder 410 according to the present invention may be configured to extract the characteristics of each atom and bond through a graph representation of the molecular structure and generate an embedding vector based on this. Alternatively, it may be configured to better understand the properties of the molecular structure by performing attention calculations related to the interactions between atoms of the plurality of molecular structures via a neural network. The decoder 420 can perform bond prediction and atom prediction, which are predicted as chemical reactions of the plurality of molecular structures, using the embedding vectors obtained by the encoder 410. The chemical reaction prediction model 400 can stabilize the chemical reaction results obtained by bond prediction and atom prediction by performing a diffusion feedback process and generate the final chemical reaction result. The configuration and detailed process of the chemical reaction prediction model 400 according to the present invention will be discussed in more detail later.

[0080] On the other hand, the chemical reaction prediction model 400 according to the present invention can utilize various chemical information such as reaction conditions, catalysts, and temperature, in addition to molecular structure information, for more accurate prediction of chemical reactions. Such chemical information may include molecular structure data (e.g., molecular structure graphs) and descriptive data (e.g., text data). Chemical reactions are influenced by various factors such as molecular structure changes, reaction mechanisms, reaction conditions, and reaction environments, so these factors must be considered comprehensively for more accurate predictions. Such chemical information is contained in various literature, textbooks, papers, patent documents, articles, and academic journals, and the chemical reaction prediction model 400 according to the present invention can make more accurate chemical reaction predictions by integrating such chemical information.

[0081] For this purpose, as shown in Figure 2b, the chemical reaction prediction model 400 may include at least one of the first module 400a and the second module 400b. Here, the first module 400a is also called the "ChemExpert-Graph" module, graph module, graph processing module, graph model, graph processing model, etc., and the second module 400b is also called the "ChemExpert-Text" module, text module, text processing module, text model, text processing model, etc.

[0082] The first module 400a can take a molecular (or chemical) structure as input and predict a chemical reaction based on a graph. For example, referring to Figure 2d(a), one embodiment of a chemical reaction can be seen. Also, as shown in Figure 2d(b), the first module 400a may include multiple layers 400a-1 to predict a chemical reaction based on a graph. More specific details about the multiple layers 400a-1 will be described later.

[0083] The molecular structure (or molecular structural formula) 403a input to the first module 400a is converted into a molecular graph, in which atoms may be represented as nodes and interatomic bonds as edges.

[0084] The molecular structure 403a may be data extracted from a document 20 containing the molecular structure 403a using the document understanding model 300, or it may correspond to at least one piece of information extracted through the storage unit 140 (or memory).

[0085] Module 400a can analyze changes in the structural properties of molecules based on the input molecular graph, predict chemical reaction pathways and the products generated as a result of the chemical reactions, and output the predicted chemical reaction pathways and products.

[0086] In one embodiment, the first module 400a can analyze structural changes of a molecule based on a molecular graph and predict the process by which specific bonds are separated and new bonds are formed.

[0087] In other embodiments, the first module 400a can analyze the interactions between atoms within a molecule based on a molecular graph and predict the formation of radicals and binding changes that may occur at each step.

[0088] In other words, the first module 400a may be configured to accept a molecular graph as input and output a predicted chemical reaction pathway and products based on the molecular graph.

[0089] Next, the second module 400b may be configured to process text data 403b and understand and predict the reaction mechanism. In this case, the text data 403b may be data extracted from a document containing molecular structure 403a using the document understanding model 300, or it may correspond to at least one piece of information extracted through the storage unit 140 (or memory) related to molecular structure 403a. The second module 400b may be a model that has been pre-trained on a vast amount of data related to chemical reactions.

[0090] In one embodiment, the text data 403b input to the second module 400b is data containing a description of the molecular structure 403a, and may include at least one of the following: chemical reaction conditions, chemical reaction mechanism (or reaction pathway), and chemical properties of the molecular structure 403a.

[0091] The second module 400b can analyze the input text data 403b to understand and predict the chemical reaction mechanism. More specifically, the second module 400b can analyze the input text data 403b and output at least one of the following: chemical reaction conditions, chemical reaction mechanism, or chemical properties, based on the text data 403b.

[0092] In one embodiment, the second module 400b can analyze the text data 403b using natural language processing (NLP) techniques and extract at least one text from among the chemical reaction conditions, chemical reaction mechanism (or reaction pathway), chemical properties, and experimental data contained in the text data 403b.

[0093] In another embodiment, the second module 400b can predict the chemical reaction mechanism (e.g., how a particular catalyst or condition affects the reaction) based on the text extracted through the analysis of the text data 403b, and output the predicted chemical reaction mechanism and chemical properties.

[0094] The second module 400b can analyze information from text data regarding the prediction of chemical reactions related to the multiple molecular structures.

[0095] The chemical reaction prediction model 400 can combine the output data 404a from the first module 400a and the output data 404b from the second module 400b to output the final chemical reaction prediction result (e.g., products, chemical reaction pathway, chemical reaction mechanism, etc.).

[0096] In one embodiment, as shown in Figures 2c(a) to (c), the chemical reaction prediction model 400 can generate electron flow, reaction conditions, and structural effects of a molecular structure (or chemical structure) using output data output from the first module 400a and the second module 400b. In this case, electron flow, reaction conditions, and structural effects can be represented together as graphs and text, and a molecular structure that reflects the position before and after electron movement can be generated, or the molecular structure of the product generated according to the reaction conditions can be generated.

[0097] In other words, the chemical reaction prediction model 400, by combining output data 404a and 404b output from the first module 400a and the second module 400b, respectively, can make more accurate predictions than by using a single data source alone, and allows users to intuitively recognize various factors related to chemical reactions.

[0098] Furthermore, the chemical reaction prediction model 400 can verify the chemical reaction results predicted through the first module 400a using the output data analyzed in the second module 400b. Specifically, the second module 400b can obtain at least one of the chemical reaction conditions, chemical reaction mechanism, or chemical properties analyzed based on the text data 403b. Based on the data analyzed in the second module 400b, the chemical reaction prediction model 400 can verify whether the chemical reaction results predicted and obtained in the first module 400a are consistent with experimental data or with theoretical expectations.

[0099] On the other hand, the chemical reaction prediction model 400 according to the present invention is characterized by performing chemical transformations based on a molecular graph, and for understanding this, we will briefly examine the configuration of the molecular structure graph.

[0100] Figure 2e(a) shows nodes and edges. In a molecular graph, atoms may be represented as nodes and interatomic bonds as edges. For example, as shown in Figure 2e(b), a water molecule has the molecular formula "H2O," and in this case, the atoms are H, H, and O. In this case, there are three nodes, and as shown in Figure 2e(c), the atoms H, H, and O may be represented as nodes n1, n2, and n3, respectively. There are also two interatomic bonds, OH and OH, and these bonds may be represented as edges e1 and e2, as shown in Figure 2e(c). When converting a molecular structure to a molecular graph in this way, the unique positional and phase relationships of the molecular structure can be preserved, enabling more accurate predictions. In this invention, the molecular structure to be predicted is converted to a molecular graph and input to the encoder 410 to predict a chemical reaction.

[0101] On the other hand, in the present invention, the encoder 410 can embed molecular structures into vectors using molecular graphs. Briefly considering the vectors, as shown in Figure 2f(a), the ethanol (C2H5OH) molecule can be represented by vectors having specific dimensions, as shown in Figure 2f(b). For example, vectors corresponding to each atom of ethanol can be represented as shown. The dimensions of the vectors and the information they contain can be set in various ways. For example, the diagram shows five-dimensional vectors, and each vector may include atomic type, bonding information, charge information, hybridization information, directional information, etc. Although not shown, the encoder 410 can be used for analysis by embedding vectors corresponding to the bonding information of molecules.

[0102] Below, we will more specifically examine the method for predicting chemical reactions based on molecular graphs and obtaining the resulting chemical reaction products in the chemical reaction prediction model 400 according to the present invention. The following explanation can be performed using the first module 400a discussed in Figure 2b.

[0103] Chemical reaction prediction and organic synthesis are important challenges in new drug development and / or materials science. Predicting the outcome of chemical reactions is crucial for designing new molecules and can significantly contribute to shortening product development cycles in various industrial fields. Traditionally, sequence-based models using SMILES strings have been utilized to predict chemical reactions and / or retrosynthetic pathways. SMILES strings encode chemical structures (or molecules or molecular structures) into ASCII strings, allowing them to be processed as text data using natural language processing techniques.

[0104] However, because SMILES strings do not match the natural graph representation of molecules, they have low learning efficiency and can produce chemically invalid transformations (or deformations). In particular, in the case of chemical reactions, transformations in SMILES space may not directly correspond to valid molecular graphs, potentially resulting in chemically unfeasible outputs. In other words, there is a limitation in that such transformations do not always lead to valid deformations in graph space, which is a more natural representation of molecules.

[0105] To address these problems, various graph-based (or central) approaches have been proposed to represent molecules as graphs. Graph representations can accurately depict molecular structures, enhance the interpretability of chemical reaction mechanisms, and facilitate more accurate chemical reaction predictions based on electron flow theory.

[0106] Therefore, the present invention proposes a chemical reaction prediction model based on electron flow using graph diffusion, which allows for a deep understanding of chemical reaction mechanisms and accurate prediction of chemical reaction results by modeling electron transfer through the graph structure of molecules. The prediction method according to the present invention is also called the Electron-Flow Inspired Graph Diffusion Model for Interpretable Forward Reaction Prediction method.

[0107] The chemical reaction prediction model described in this invention utilizes graph diffusion, so that both inputs and outputs operate in graph space, which can provide the user with a more interpretable transformation while maintaining invariance with respect to the permutations and orders of SMILES strings. The chemical reaction prediction model according to this invention can learn the transformation process from an initial set of graphs to the final product, rather than starting with random graph sampling.

[0108] The present invention will be discussed in more detail below, along with the attached drawings.

[0109] As shown in Figure 4a, the chemical reaction prediction model 400 according to the present invention may include an encoder 410, a decoder 420, a graph diffusion module 430, and a learning unit 440. As shown in the figure, the chemical reaction prediction model 400 can work with memory and at least one processor to perform a series of data processing according to the present invention. The memory can store various information. For example, the memory may store various information such as information input to the chemical reaction prediction model 400, information about results, intermediate products, etc., generated by a series of data processing in the chemical reaction prediction model 400, and information about the final result. The processor can work with each component of the chemical reaction prediction model to perform the chemical reaction prediction process according to the present invention in each component and neural network layer.

[0110] In the chemical reaction prediction model 400 according to the present invention, a learning unit 440 and a ground-truth product are present during the learning step, as shown in Figures 4a and 4b. However, after learning is completed, the learning unit and the ground-truth product may be removed, as shown in Figures 4c and 4d.

[0111] The encoder 410 according to the present invention may be configured as a backbone encoder. The encoder 410 according to the present invention is configured as an attention-based graph neural network (GNN) and can perform relative position encoding based on the shortest path between nodes (atoms). On the other hand, the diffusion process described in the present invention does not depend on a backbone network and can be realized by various encoders.

[0112] The encoder 410 according to the present invention may include an embedding layer 411 and a graph neural network (GNN) 412, as shown in the figure. The graph neural network may be an attention-based graph neural network or a backbone neural network.

[0113] On the other hand, the graph neural network 412 may include a multi-head self-attention layer 413 and a feed-forward neural network layer. The multi-head self-attention layer is also called a topologically weighted multi-head self-attention layer.

[0114] The embedding layer 411 is configured to receive information about multiple molecular structures and convert it into embedding vectors. The embedding layer 411 may also be configured to generate vectors that include atoms of the molecular structures and the bonding relationships between atoms. The graph neural network 412 can receive the embedding vectors embedded in the embedding layer 411, perform attention operations related to the interactions between atoms of the multiple molecular structures, and update the embedding vectors based on the operations. At this time, the embedding layer can receive a molecular graph and obtain embedding vectors based on it.

[0115] On the other hand, the node features included in the embedding vector output by the embedding layer 411 may include the atomic type, atomic charge, implicit number of hydrogen atoms, and number of radical electrons. The embedding layer 411 can encode nodes following the molecular graph and generate an embedding vector containing information that can be used within the encoder. This allows the molecular graph to be completely reconstructed for each molecule, and the node order is also tokenized as an input feature that can later be used to measure the consistency between bond predictions and predicted atomic orders.

[0116] The characteristics of atom v may also be represented as (f), (g), and (h) in Figure 5a, and may be tokenized and embedded via the embedding layer 411.

[0117] As shown in Figure 5a(g), the features of atom v are tokenized and embedded via the embedding layer 411, and the tokenized features of type i shown in Figure 5a(f) may have features as shown in Figure 5a(g). i is an index representing the type of atomic feature and can belong to the set of type, charge, hydrogen, radical, or order. i means at least one of the features of type, charge, hydrogen, radical, or order, where type may mean atomic type (e.g., carbon, oxygen, etc.), charge may mean the formal charge of the atom, hydrogen may mean the number of implicit hydrogens, radical may mean the number of radical electrons of the atom, and order may mean the order of the node (the number of other atoms bonded to that atom). Such features are embedded via the embedding layer 411 as shown in Figure 5a(h), forming the initial embedding of the atom. The initial embedding of atom v may be the average of all embeddings.

[0118] The embedding layer 411 can perform embedding for multiple molecules based on a molecular graph. The molecular graph relating to molecular structure may exist stored in memory, a dataset, a database, etc. Alternatively, the molecular graph relating to molecular structure may be extracted by the molecular graph extraction module 450. The chemical reaction prediction model 400 may further include the molecular graph extraction module 450 and cooperate with the molecular graph extraction module 450. Furthermore, it goes without saying that the molecular graph extraction module 450 may be included as a component of the encoder 410.

[0119] Molecular graphs can maintain multiple molecular structures by representing atoms as nodes and the bonds between atoms as edges.

[0120] Multiple molecular structures may include a first molecular structure and a second molecular structure, and such first and second molecular structures are also called reactants. These correspond to reference numerals 401a and 401b in Figure 4a.

[0121] The molecular graph extraction module 450 can obtain molecular graphs relating to the multiple molecular structures by using a pre-identified graph transformation algorithm to transform atoms into nodes and interatomic bonds into edges, based on the multiple molecular structures.

[0122] The molecular graph extraction module 450 can obtain a first molecular graph including nodes and edges corresponding to the first molecular structure by converting the atoms constituting the first molecular structure into nodes and the bonding relationships between the atoms constituting the first molecular structure into edges. Similarly, the molecular graph extraction module 450 can obtain a second molecular graph including nodes and edges corresponding to the second molecular structure by converting the atoms constituting the second molecular structure into nodes and the bonding relationships between the atoms constituting the second molecular structure into edges using the pre-specified graph transformation algorithm.

[0123] In the embedding layer 411 of the encoder 410, embedding vectors corresponding to the plurality of molecular structures can be obtained using the information about the plurality of molecular structures. In this case, the information about the plurality of molecular structures may include information about nodes and edges corresponding to the first molecular structure and information about nodes and edges corresponding to the second molecular structure. The information about molecular structures may be a molecular graph extracted by the molecular graph extraction module 450, or information extracted from such a molecular graph.

[0124] The embedding vector may include at least one piece of information regarding the atomic type, atomic charge, number of hydrogen atoms, and the number and degree of radical electrons of the nodes corresponding to the first molecular structure and the nodes corresponding to the second molecular structure. Furthermore, the embedding vector may further include information regarding the bonding relationships of the atoms in the first and second molecular structures.

[0125] On the other hand, once the reactants to be predicted, i.e., multiple molecular structures (for example, a first molecular structure 401a and a second molecular structure 401b), are identified, the chemical reaction prediction model 400 can use the nodes and edges corresponding to the first molecular graph and the nodes and edges corresponding to the second molecular graph to extract at least one of the following: an adjacency matrix, a bond type matrix, shortest paths, and K-hop neighbors, respectively, for the first and second molecular graphs. The timing at which such information 480 is extracted is not limited and may vary.

[0126] The adjacency matrix may also include information regarding direct connections between nodes constituting the first molecular graph and the second molecular graph.

[0127] The bond type matrix may include information regarding the bond types between nodes constituting the first molecular graph and the second molecular graph.

[0128] The shortest path matrix may include information regarding the shortest path lengths between nodes constituting the first molecular graph and the second molecular graph.

[0129] The K-hop neighborhood information may include information about neighboring nodes within K hops for each node constituting the first molecular graph and the second molecular graph.

[0130] At least one of the extracted information 480, consisting of an adjacency matrix, bond type matrix, shortest paths, and K-hop neighbors, may be stored in memory. At least one of the adjacency matrix, bond type matrix, shortest paths, and K-hop neighbors stored in memory may be input to at least one of the embedding layer 411, graph neural network 412, multi-head self-attention layer 413, and feedforward neural network layer 414.

[0131] In the embedding layer 411, when embedding vectors v1 corresponding to multiple molecular structures (or reactants, e.g., first molecular structure 401a, second molecular structure 401b) are obtained, the obtained embedding vectors v1 may be input to the graph neural network 412. In particular, the embedding vectors v1 may be input to the multi-head self-attention layer 413 of the graph neural network.

[0132] The Multi-Head Self-Attention layer 413 can perform attention calculations related to the interactions between atoms of multiple molecular structures using an embedding vector v1 embedded based on a molecular graph, and update the embedding vector based on the calculation. In this case, the embedding vector v1 to be updated may be a vector output from the embedding layer.

[0133] The multi-head self-attention layer 413 can perform global attention-based pooling. The multi-head self-attention layer 413 can add a bias to the number of attention weight points derived from the shortest paths between node pairs to account for molecular graph topology. This can act as a relative position encoding between atoms. Consequently, this is unaffected by spectrum-based encoding problems (e.g., lack of sign invariance for Laplace matrix eigenvectors) because the distance measurements between nodes are not absolute. Such encoding is computationally efficient and allows for the consideration of the entire graph topology while allowing for the transmission of global messages. The timing of the shortest path calculations may vary.

[0134] Global attention-based pooling considers the entire topology of the graph, and the multi-head self-attention layer 413 can globally consider the relationships between each node in the graph by performing global attention-based pooling.

[0135] For example, assuming the input vector from the embedding layer has a dimension of 128, and the number of heads in the multi-head attention system is 8, each head may be input a 16-dimensional sub-vector, which is the 128-dimensional input vector divided by the number of heads, 8.

[0136] Each head of the multi-head self-attention layer 413 is configured to independently calculate attention values ​​for 16-dimensional subvectors by applying a pre-configured attention mechanism. For example, the first head may primarily reflect information about adjacent nodes in the molecular graph, and the second head may primarily reflect information about a specific bond type in the molecular graph. The 16-dimensional attention vectors calculated by each of the multiple heads of the multi-head self-attention layer 413, for example, eight heads, may be combined again to form a 128-dimensional vector. This combined vector has the same dimensions as the embedding vector output from the embedding layer 411, but can contain richer information by integrating various information extracted from each head. In this invention, this can be expressed as an updated embedding vector.

[0137] On the other hand, in order to reflect the shortest path between nodes in the weights, the present invention can calculate the shortest path between all pairs of nodes in the molecular graph, as described above. The shortest path can be calculated by a pre-set algorithm. The multi-head self-attention layer 413 can encode the relative positions between nodes using the calculated shortest path information. This encoding can be done based on the shortest path distance between nodes. When the multi-head self-attention layer 413 calculates the number of attention weight points between nodes, it can use the shortest path information as a bias. For example, the multi-head self-attention layer 413 can adjust the attention weights by increasing the attention weight when the shortest path distance between two nodes is close, and decreasing it when the shortest path distance between two nodes is farther. The multi-head self-attention layer 413 can add the bias derived from the shortest path between pairs of nodes to the number of attention weight points.

[0138] The multi-head self-attention layer 413 thus encodes each node of the molecular graph using attention weights adjusted based on the shortest path, which enables encoding that takes into account the entire topology of the molecular graph.

[0139] Furthermore, the multi-head self-attention layer 413 can add a bias to the attention score based on the bond type (e.g., single, double, triple, aromatic bond).

[0140] For example, in the multi-head self-attention layer 413, different biases can be applied to the attention score calculated in the multi-head self-attention layer, depending on the type of binding between the nodes constituting the first molecular graph and the second molecular graph.

[0141] Here, the bond type may include single bond type, double bond type, triple bond type, and aromatic bond type.

[0142] This may mean that instead of a single bias matrix for all heads, each head has its own unique bias matrix. For example, the first head may have a shortest path bias added, and the second head may have a single bond bias added. The remaining heads can use masks that focus on interatomic or intermolecular global attention. In the present invention, in the multi-head self-attention layer 413, such various bias matrices can be added to each layer with multiple heads. In the present invention, by adding bias matrices to each layer corresponding to multiple heads, the attention score of each head can highlight information masked by the bias matrix. For example, if a bias matrix corresponding to the bond type is added, the head can receive information only from neighbors connected by a specific bond type, masked according to the bond type. For example, a particular head can receive information only from nodes connected by single bonds.

[0143] Thus, the weights for the attention score according to the present invention, which add biases that reflect various properties of the molecular structure, can be calculated according to the formulas and meanings shown in Figure 5. Here, A(Qk, Kk, Vk) may represent the attention weights calculated by applying the softmax function to the attention score.

[0144] In this way, the multi-head self-attention layer 413 performs attention calculations related to the interactions between atoms of the plurality of molecular structures, and updates the embedding vector based on the calculations. The updated embedding vector can then be stored in memory. The updated embedding vector stored in memory can then be input to the feed-forward neural network layer 414. The feed-forward neural network layer 414 receives the output of the multi-head self-attention layer 413 and can further update the updated embedding vector. The feed-forward neural network layer 414 can update the output vector of the multi-head self-attention layer using at least one of the adjacency matrix, bond type matrix, shortest paths, and K-hop neighbors. The feed-forward neural network layer 414 can update the output vector of the multi-head self-attention layer so that it contains more structural information about the molecular structures. The output vector of the feedforward neural network layer can be identified as the final updated embedding vector v2 input to the decoder 420.

[0145] Furthermore, the decoder 420 according to the present invention may include at least one projection layer 421. The decoder 420 can perform bond prediction and atom prediction based on multiple molecular structures (reactants) based on an updated embedding vector v2 input to the projection layer. Bond prediction predicts the bonds between atoms, and atom prediction predicts the state changes of atoms. The decoder 420 may include multiple projection layers 422a, 422b. Bond prediction 423 and atom prediction 424 can be performed in each of the projection layers 422a, 422b.

[0146] In decoder 420, coupling prediction can be performed by executing a dot product operation using the updated embedding vectors. The dot product operation can be performed for each pair of atoms corresponding to the updated embedding vectors.

[0147] In block 423 in the drawing, each box on the x-axis 423a and y-axis 423b conceptually represents an atom of the molecule to be reacted, and such atoms can be represented as the updated embedding vector v2. Based on the information contained in the updated embedding vector, the decoder 420 can predict the bond type between each atom through an inner product operation.

[0148] The decoder 420 can obtain dot product values ​​for each of the multiple bond types for each of the atom pairs based on the dot product calculation. The multiple bond types can relate to at least one of single bond, double bond, bond formation, bond collapse, and no change.

[0149] The decoder 420 can generate probability distributions for each of the multiple bond types for each of the atom pairs using the dot product values ​​obtained by the dot product calculation. Furthermore, the decoder 420 can obtain a transformation matrix that predicts the change in the bond state of each of the atom pairs using the probability distributions.

[0150] The decoder 420 can generate the probability distribution of the multiple bond types for each of the atomic pairs by applying a softmax function to the dot product value obtained for each of the multiple bond types for each of the atomic pairs.

[0151] The decoder 420 can generate a probability distribution of atomic properties for each atom corresponding to the updated embedding vector v2 using a softmax output layer. The decoder 420 can also predict the atomic properties of the atoms corresponding to the updated embedding vector using the probability distribution. The atomic properties may include the charge state of the atoms, which can change during the chemical reaction process of the plurality of molecular structures.

[0152] The decoder 420 can perform the dot product between the updated embedding vectors v2 (or atomic embedding vectors, shown as in Figure 5b(a)) for the atoms and calculate the number of potential bond transformation points, as shown in Figure 5(b). Here, the lowercase t can represent a stage in the diffusion process.

[0153] This is based on the standard query-key matrix multiplication of the transformer attention mechanism, but in this invention, the key and query can share the same projection head. The setting that the key and query have the same projection head may require many projections to accommodate various bond types such as none, single, double, triple, and aromatic. This corresponds to multi-head attention in the transformer, which can be expressed mathematically as shown in Figure 5b(c). Thus, bond prediction can be performed in bond prediction heads, which may correspond to the first projection layer 422a shown in the figure. That is, bond prediction can be performed in the first projection layer 422a. The first projection layer 422a may consist of layers that make up the bond prediction head.

[0154] Each bond type prediction can be considered a head in this multi-head attention scheme, where softmax normalization is applied to a state like that shown in Figure 5b(d) to obtain the distribution of bond transformations for each atom pair, where Figure 5b(d) may represent a bond transformation.

[0155] The join prediction head calculates a score S for each join type k, as shown in Figure 5b(c), where the meaning of each symbol is as shown in Figure 5b(d).

[0156] On the other hand, the join prediction head can calculate the probability of a join transformation T for a join type k, as shown in Figure 5b(e), as shown in Figure 5b(b). The join prediction head can obtain the probability distribution of the join transformation by applying softmax normalization as shown in Figure 5b(f).

[0157] In this invention, each bond type prediction can be understood as each head in a multi-head attention system. The decoder 420 is applied to the bond transformation-possible states shown in Figure 5b(e) through softmax normalization to obtain a normalized probability distribution of the bond transformation for each atom pair. This can provide an efficient and scalable method for directly predicting bond transformations in the embedding space.

[0158] Next, atomic prediction can be performed in the second projection layer 423b of the decoder 420. The second projection layer 423b may consist of layers that constitute an atomic prediction head. Since the characteristics of some atoms may change during a chemical reaction, the decoder 420 according to the present invention can not only predict bond transformations between atoms but also make predictions about atoms. Atomic prediction can be performed in the second projection layer 423b. For this purpose, a standard softmax output layer can be used for each atomic characteristic. This includes the order of the atom for each bond type (e.g., the number of single bonds the atom has), and such predictions can contribute to ensuring consistency between independent parallel samples of bond transformations and atomic transformations.

[0159] The atomic prediction head can calculate the probability that the feature Av of atom v becomes a at step t of the diffusion process, as shown in Figure 5c(c). Here, Figure 5c(b) represents the embedding of atom v, and Figure 5c(d) can represent the weight matrix of the prediction head for atom feature a. The atomic prediction head can calculate the probability that the atom feature Av becomes a by multiplying the embedding h of atom v (see Figure 5c(b)) by the weight matrix W (see Figure 5c(d)) and then applying the softmax function.

[0160] The decoder 420 can obtain predicted chemical reaction results (probability distributions) 425 from chemical reactions of multiple molecular structures by combining the predicted atomic bonding probabilities and atomic prediction probabilities output from the projection layers 422a and 422b. The chemical reaction results predicted by the projection layers are also called initialization reaction results.

[0161] In the present invention, the predicted chemical reaction results may be input to a Graph Diffusion Module 430, which can sample the initialization reaction results using the bond prediction results and the atom prediction results. The Graph Diffusion Module 430 can stabilize the sampled initialization reaction results through a diffusion feedback process. The chemical reaction prediction model can obtain the final chemical reaction results stabilized through the diffusion feedback process.

[0162] In the diffusion feedback process, the graph diffusion module 430 can repeatedly evaluate each bond transformation of the initialization reaction result and remove or modify unstable bonds. Furthermore, in order to predict changes in the bond state of the initialization reaction result, the graph diffusion module 430 can evaluate the predicted bond transformations at each of the repeatedly executed steps using a transformation probability matrix and a target transformation matrix, and generate the final chemical reaction result using an interpolation factor.

[0163] In this invention, since bond prediction is performed independently across the entire molecular graph, a method is needed to obtain a valid molecular structure in the end. For example, the prediction process may result in the prediction of bonds to multiple atoms or violations of bonding rules. To solve such problems, a denoising process is needed to obtain a valid final state.

[0164] As shown in Figure 5e, the predicted chemical reaction result can be stabilized by performing multiple sampling and denoising processes 432.

[0165] Figure 5e shows the diffusion process from the ground truth data 440 (or target label, ground truth product) to the initial distribution 425, which is based on an initial distribution derived from the dot product matrix of bond transformation predictions between all atomic pairs. Here, the target label corresponding to the ground truth data may include the actual bond states and atomic states corresponding to multiple molecular structures input to the encoder. Here, the initial distribution 425 refers to the initialization chemical reaction result 425 output from the decoder, which may correspond to a probability distribution. On the other hand, the diffusion process to the target label takes place during the learning process in the learning unit 440, and in the actual inference stage, a fully trained graph diffusion module 430 may be used. In this case, the predicted chemical reaction result 425 is input to the fully trained graph diffusion module 430, and after a denoising process, the final chemical reaction result can be obtained.

[0166] The learning unit 440 calculates a loss function between the final chemical reaction result output from the decoder and label data (hereinafter also called "target label") which includes the actual bonding and atomic states corresponding to the plurality of molecular structures, and can optimize at least one of the parameters of the encoder and the decoder in order to minimize the loss function.

[0167] On the other hand, during the learning process, the target label may be interpolated with the initial distribution and sampled at random points in the interpolated space. Figure 5e shows a noise graph in which such random points are denoised to the target label. During prediction, the results of sampling with the initial distribution are updated at each step to obtain a stable final product molecule, and a similar process can be applied to atomic features.

[0168] In the Diffusion Step, the Predicted Product, as described above, may correspond to the initial distribution obtained by the decoder 420 via the dot product matrix of bond transformation predictions for all atomic pairs. This initial distribution represents the probability distribution of bond formation between atoms. The Diffused Product is obtained by interpolating the ground-truth product with the initial distribution, and in this invention, a noisy graph can be obtained by sampling random points in this interpolated space. Such random points represent a noise graph that is denoised towards the target label. The Denoised Product is obtained by denoising the sampled noise graph to gradually approach the target label, and this process is carried out in multiple stages, with the graph changing more stably at each stage. The Alternative Product means that, in some cases, an alternative product may be obtained through a different path, which may have a different bond configuration than the target label.

[0169] Thus, in this invention, the decoder 420 acquires an initial distribution, and the spread graph module 430 samples the initial distribution to sample random points. The spread graph module 430 interpolates the target label with the initial distribution through interpolation and denoising processes to obtain an intermediate state. During the learning process, such an intermediate state v3 may be input again to the encoder. The spread graph module 430 samples random points to obtain a graph containing noise and gradually approaches the target label through a denoising process. Such a process may be performed repeatedly, and in such iterative processes, at least one of the encoder, decoder, and spread graph module can be learned.

[0170] The graph diffusion module 430 may be configured to learn how to denoise the predicted chemical reaction results 425 by first spreading the target label into the previous output distribution, and then learning how to map the states in the diffused interpolation space to the final target label.

[0171] For this purpose, a transformation probability matrix Mt and a desired target transformation matrix T can be defined, and the transformation matrix T can be derived from the atomic bond prediction results. The atomic bond prediction results are obtained from the decoder and can be derived from the bond prediction head described above. This can be represented as the atom-mapped label of the reaction. The transformation matrix T can identify whether bonds are formed or broken by calculating the difference between the input and output adjacency matrices, and the transformation matrix T can be represented as shown in Figure 5d(a). This defines the difference between the adjacency matrices of the reactants and products, and the adjacency matrix can represent the entire bond type matrix including single / double / triple / aromatic bonds.

[0172] On the other hand, the reactants change continuously at each diffusion step and are therefore time-dependent. This means that the target label also changes at each step, and in this invention, a method can be learned to predict when and where the difference between the reactants and the target label occurred (reaction center), and if a difference occurred, what changed (final combined configuration), as shown in Figure 5d(b). The parameter at is a randomly sampled interpolation factor that determines the weights of the current transformation matrix and the target transformation matrix. During learning, the diffusion process may include obtaining the transformation probability Mt from the initial reactant set {Gr} and sampling {Gr(t+1)} at Mt+1 to use as input for the next step. Either step may aim to directly predict the target transformation matrix T. Therefore, the learning loss can be defined as shown in Figure 5d(c). During prediction, the input {Gr} can simply be encoded, the probability of the combined transformation Mt can be obtained using the prediction head, and sampling can be performed at each step using standard categorical softmax sampling on this distribution. In the learning process of this invention, this process can be continuously repeated and the results fed back until a stable configuration is obtained in which the combinations no longer change.

[0173] On the other hand, in this invention, the final chemical reaction result can be derived by applying Top-K ranking to multiple results stabilized in the diffusion module 430. This may be performed in the Top-K Sampled Product module of the chemical reaction prediction model 400. The Top-K ranking process may be performed in the inference stage, that is, it can be used to extract the results with the highest probability in order to provide the user with the final result. In this invention, at least one of Top-K sampling and Top-K ranking can be performed, which can be used to obtain a wider variety of results in the inference process. In Top-K sampling, multiple samples can be generated based on what the chemical reaction prediction model has learned, and the distribution can be diversified by adjusting the temperature parameter τ. Top-K ranking is used to select the most likely result from the generated samples, the probability of each sample can be calculated, and the result with the highest probability can be selected as the final prediction.

[0174] In the Top-K sampling process, when sampling the distribution Mt at a time step t, the logits can be readjusted by the temperature parameter τ. This allows for sharper or more uniform adjustment of the softmax distribution shown in Figure 6(a).

[0175] On the other hand, in order to obtain multiple samples, the present invention can use a temperature-mediated parameter (temperature schedule) τ. In the present invention, when attempting to extract N samples, τi can be set as shown in Figure 6(b). Here, f(i) is the temperature sampling schedule, and in the present invention, f(i) can be defined as shown in Figure 6(c). Here, x is the linspace between 0 and 1.

[0176] Next, we consider the Top-K ranking. When sampled candidates are available, a method is needed to rank them based on probability. This can be done by calculating probabilities using the transformation matrix Mt. If the sampled states are considered as sampled tokens for each atom and bond, the probability of each state can be derived from the sampled softmax distribution Mt. This allows us to assign a total probability to the final overall configuration of the molecule (atomic and bond states) and calculate the sampled probabilities (S). This can be shown as in Figure 6(d). Here, Figures 6(e), (f), and (g) can be derived from the softmax distribution for atomic features, bond features, and atomic order, respectively. Furthermore, term 6(d) can ensure consistency between the predicted atomic order and the actual order calculated from the bonds. In this invention, since bonds are sampled independently, multiple bonds can be formed on the same atom. Therefore, the actual order of each atom in the sampled graph is directly calculated and compared with the predicted probability distribution of the order. In this method, if the actual order of the generated molecule matches the predicted order, a high probability score is assigned.

[0177] Furthermore, since there is no guarantee that all solutions are unique when sampling independently of the distribution, many duplicate samples may occur. Therefore, in this invention, the Top-K products can be selected during prediction by taking a simple, unique graph.

[0178] On the other hand, the above explanation focused on Figures 4a and 4b, which both include the learning process of the chemical reaction prediction model. However, as mentioned above, in the inference process, the chemical reaction prediction model can have the configurations shown in Figures 4c and 4d. In this case, if the learning process is excluded, everything can be understood similarly, so please refer to the explanation above for a detailed explanation.

[0179] In the inference process, as described above and as shown in Figure 3, the encoder receives information about a plurality of molecular structures, including a first molecular structure and a second molecular structure, as input (S310); the embedding layer of the encoder uses the information about the plurality of molecular structures to obtain embedding vectors corresponding to the first molecular structure and the second molecular structure (S320); the multi-head self-attention layer performs attention calculations related to the interaction between atoms of the first molecular structure and the second molecular structure and updates the embedding vectors based on the calculations (S330); the updated embedding vectors are stored in memory and input to the decoder (S340); and the decoder uses the updated embedding vectors to perform bond prediction and atom prediction, which are predicted as chemical reactions of the first molecular structure and the second molecular structure. The final chemical reaction product can be obtained through a process of performing prediction (S350) and a process of obtaining the final chemical reaction product predicted from the chemical reactions of the first molecular structure and the second molecular structure using the results of the bond prediction and the atom prediction (S360).

[0180] The process for obtaining such a final chemical reaction result can be carried out through the response generation system 100 discussed in Figure 1. In this system, the response generation system 100 can receive user queries, and these user queries may include information about the molecular structure to be predicted for the chemical reaction.

[0181] In the chemical reaction prediction model 400 according to the present invention, as shown in Figure 7(a), the chemical reaction result can be obtained through bond prediction and atom prediction for multiple molecular structures that are the target of the chemical reaction, as shown in Figure 7(b).

[0182] As described above, the chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system according to the present invention allow for a deep understanding of the chemical reaction mechanism and accurate prediction of the chemical reaction results by modeling electron transfer through the graph structure of molecules.

[0183] Furthermore, the chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system according to the present invention provide a chemical reaction prediction model that can deeply understand the chemical reaction mechanism and accurately predict the chemical reaction results, thereby reducing the time and cost of experiments for users. This reduces research and development costs and shortens the time to market for new products.

[0184] On the other hand, according to the chemical reaction prediction system, its control method, and learning method for the chemical reaction prediction system of the present invention, by performing both the input of molecular structures and the output for predicted results in graph space, forward reaction prediction can be performed regardless of the permutation and order of SMILES.

[0185] Furthermore, according to the chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system of the present invention, it is possible to ensure consistency between transformations by enabling simultaneous sampling of multiple interdependent transformations that occur in parallel within the molecular graph.

[0186] Furthermore, according to the chemical reaction prediction system, its control method, and the learning method for the chemical reaction prediction system of the present invention, in order to solve problems that may arise due to symmetrical structures, it is possible to break the symmetry by including noise or a sampling mechanism and form an effective output structure, thereby preventing the occurrence of ineffective configurations.

[0187] On the other hand, the present invention described above can be embodied as a program that is executed by one or more processes on a computer and can be stored on a medium (or recording medium) that is readable by such a computer.

[0188] Furthermore, the present invention described above can be embodied as computer-readable code or instruction words on a medium on which a program is recorded. That is, the present invention can be provided in the form of a program.

[0189] On the other hand, computer-readable media include all types of recording devices that store data readable by a computer system. Examples of computer-readable media include HDDs (Hard Disk Drives), SSDs (Solid State Disks), SSDs (Silicon Disk Drives), ROMs, RAMs, CD-ROMs, magnetic tapes, floppy disks, and optical data storage devices.

[0190] Furthermore, the computer-readable medium may include storage and may be a server or cloud storage accessible by electronic devices via communication. In this case, the computer can download the program according to the present invention from the server or cloud storage via wired or wireless communication.

[0191] Furthermore, in this invention, the computer described above is an electronic device equipped with a processor, i.e., a CPU (Central Processing Unit), and its type is not particularly limited.

[0192] On the other hand, the above detailed description should not be interpreted restrictively in any way, but should be considered illustrative. The scope of the invention should be determined by a reasonable interpretation of the appended claims, and all modifications within the scope of the equivalents of the invention are included within the scope of the invention.

Claims

1. A method for predicting chemical reactions, which is performed in cooperation with memory and a processor, The process involves receiving information about multiple molecular structures as input to an encoder, The embedding layer of the encoder includes the step of obtaining an embedding vector corresponding to the plurality of molecular structures using information about the plurality of molecular structures, The steps include: performing attention calculations related to the interactions between atoms of the multiple molecular structures in a multi-head self-attention layer, and updating the embedding vectors based on the calculations; The steps include storing the updated embedding vector in memory and inputting the updated embedding vector stored in memory to the decoder, The decoder performs bond prediction and atom prediction, which are predicted as chemical reactions of the plurality of molecular structures, using the updated embedding vector. A method for predicting a chemical reaction, comprising the step of obtaining a final chemical reaction result predicted from the chemical reaction of the plurality of molecular structures using the results of the bond prediction and the results of the atom prediction.

2. The step of obtaining the chemical reaction result is: The steps include sampling the results of the initialization reaction using the bond prediction results and the atom prediction results, The steps include stabilizing the sampled initialization reaction result through a diffusion feedback process, A method for predicting a chemical reaction according to claim 1, characterized by comprising the step of obtaining the final chemical reaction result stabilized through the diffusion feedback process.

3. The method further includes the step of obtaining a molecular graph relating to the plurality of molecular structures by converting atoms to nodes and interatomic bonds to edges, based on the plurality of molecular structures, The aforementioned plurality of molecular structures include a first molecular structure and a second molecular structure, The step of obtaining the aforementioned molecular graph is: A first molecular graph including nodes and edges corresponding to the first molecular structure is obtained by converting the atoms constituting the first molecular structure into nodes and the bonding relationships between the atoms constituting the first molecular structure into edges, using a pre-specified graph transformation algorithm. The chemical reaction prediction method according to claim 1, further comprising the step of obtaining a second molecular graph including nodes and edges corresponding to the second molecular structure by converting atoms constituting the second molecular structure into nodes and bonding relationships between atoms constituting the second molecular structure into edges using the aforementioned predetermined graph conversion algorithm.

4. The information regarding the aforementioned multiple molecular structures is, This includes information relating to nodes and edges corresponding to the first molecular structure, and information relating to nodes and edges corresponding to the second molecular structure, The aforementioned embedding vector is The chemical reaction prediction method according to claim 3, characterized in that it includes at least one piece of information relating to the atomic type, atomic charge, number of hydrogen atoms, and the number and degree of radical electrons of the node corresponding to the first molecular structure and the node corresponding to the second molecular structure.

5. In the step of updating the aforementioned embedding vector, The chemical reaction prediction method according to claim 4, characterized in that different biases are applied to the attention score calculated in the multi-head self-attention layer, according to the binding types between the nodes constituting the first molecular graph and the second molecular graph.

6. The chemical reaction prediction method according to claim 5, characterized in that the bond type includes single bond type, double bond type, triple bond type, and aromatic bond type.

7. Using the nodes and edges corresponding to the first molecular graph and the nodes and edges corresponding to the second molecular graph, The process further includes the step of extracting at least one of the following: an adjacency matrix, a bond type matrix, a shortest path matrix, and K-hop neighbors, corresponding to the first molecular graph and the second molecular graph, respectively. The aforementioned adjacency matrix is, This includes information regarding direct connections between nodes constituting the first molecular graph and the second molecular graph, The aforementioned join type matrix is, This includes information regarding the types of connections between nodes constituting the first molecular graph and the second molecular graph, The aforementioned shortest path matrix is, This includes information regarding the shortest path length between nodes constituting the first molecular graph and the second molecular graph, The aforementioned K-hop proximity information is, The chemical reaction prediction method according to claim 5, characterized in that each node constituting the first molecular graph and the second molecular graph includes information about neighboring nodes within K hops.

8. In the step of updating the aforementioned embedding vector, The output vector of the multi-head self-attention layer is input to a feedforward neural network layer. In the feedforward neural network layer, the output vector of the multi-head self-attention layer is updated using at least one of the adjacency matrix, bond type matrix, shortest path matrix, and K-hop neighbor information. The chemical reaction prediction method according to claim 7, characterized in that the output vector of the feedforward neural network layer is identified as the updated embedding vector.

9. The aforementioned combination prediction is performed by executing a dot product operation using the updated embedding vector. The chemical reaction prediction method according to claim 1, characterized in that the dot product calculation is performed for each pair of atoms corresponding to the updated embedding vector.

10. The step of performing the aforementioned coupling prediction is: The step includes obtaining the dot product value for each of the multiple bond types for each of the atomic pairs based on the dot product calculation, The method for predicting a chemical reaction according to claim 9, characterized in that the plurality of bond types relate to at least one of single bond, double bond, bond formation, bond collapse, and no change.

11. The step of performing the aforementioned coupling prediction is: The steps include generating a probability distribution for each of the plurality of bond types for each of the atomic pairs using the dot product value obtained by the dot product calculation, The method for predicting a chemical reaction according to claim 10, further comprising the step of obtaining a transformation matrix that predicts the change in the bonding state of each atomic pair using the probability distribution.

12. In the step of generating the aforementioned probability distribution, A method for predicting a chemical reaction according to claim 11, characterized in that a softmax function is applied to the dot product value obtained for each of the plurality of bond types for each of the atom pairs to generate the probability distribution of the plurality of bond types for each of the atom pairs.

13. In the step of performing the aforementioned atomic prediction, A step of generating an atomic characteristic probability distribution for each atom corresponding to the updated embedding vector using a softmax output layer, The method further includes the step of predicting the atomic properties of the atoms corresponding to the updated embedding vector using the probability distribution, The chemical reaction prediction method according to claim 1, characterized in that the atomic properties include the charge state of the atoms which can change during the chemical reaction process of the plurality of molecular structures.

14. The aforementioned diffusion feedback process is, The chemical reaction prediction method according to claim 2, characterized by repeatedly evaluating each bond transformation of the initial chemical reaction result and removing or changing unstable bonds.

15. The aforementioned diffusion feedback process is, A method for predicting a chemical reaction according to claim 14, characterized in that, in order to predict the change in the binding state of the initialization reaction result, the predicted binding transformation is evaluated for each of the repeatedly performed steps using a transformation probability matrix and a target transformation matrix, and the final chemical reaction result is generated using an interpolation factor.

16. A method for predicting chemical reactions, which is performed in cooperation with memory and a processor, The process involves receiving information about multiple molecular structures as input to an encoder, A step of obtaining a molecular graph using atoms as nodes and bonds as edges based on the aforementioned multiple molecular structures, The embedding layer of the encoder includes the step of obtaining an embedding vector corresponding to the molecular graph using the information on the plurality of molecular structures, The steps include: performing attention calculations related to the interactions between atoms of the multiple molecular structures in a multi-head self-attention layer, and updating the embedding vectors based on the calculations; The steps include storing the updated embedding vector in memory and inputting the updated embedding vector stored in memory to the decoder, The decoder performs bond prediction and atom prediction, which are predicted as chemical reactions of the plurality of molecular structures, using the updated embedding vector. The steps include obtaining the final chemical reaction product predicted from the chemical reactions of the plurality of molecular structures using the results of the bond prediction and the results of the atom prediction, The steps include: calculating a loss function between the final chemical reaction result and label data including the actual bonding and atomic states corresponding to the plurality of molecular structures; A learning method for predicting chemical reactions, comprising the step of optimizing the parameters of at least one of the encoder and the decoder in order to minimize the loss function.

17. A chemical reaction prediction system comprising memory, encoder, decoder, and at least one processor, The encoder receives information on multiple molecular structures and, based on the multiple molecular structures, obtains a molecular graph using atoms as nodes and bonds as edges. In the embedding layer of the encoder, an embedding vector corresponding to the molecular graph is obtained. In the multi-head self-attention layer of the encoder, attention calculations related to the interactions between atoms of the plurality of molecular structures are performed, and the embedding vector is updated based on the calculations. The aforementioned processor, The updated embedding vector is stored in memory in the update step, and the updated embedding vector stored in memory is input to the decoder. The decoder mentioned above is Using the updated embedding vectors, bond prediction and atom prediction are performed to predict the chemical reactions of the multiple molecular structures. A chemical reaction prediction system characterized by obtaining the final chemical reaction result predicted from the chemical reactions of the plurality of molecular structures using the results of the bond prediction and the results of the atom prediction.

18. A program executed by one or more processes in an electronic device and stored on a computer-readable medium, The aforementioned program, The process involves receiving information about multiple molecular structures as input to an encoder, A step of obtaining a molecular graph using atoms as nodes and bonds as edges based on the aforementioned multiple molecular structures, The steps include obtaining an embedding vector corresponding to the molecular graph in the embedding layer of the encoder, The steps include: performing attention calculations related to the interactions between atoms of the multiple molecular structures in a multi-head self-attention layer, and updating the embedding vectors based on the calculations; The steps include storing the updated embedding vector in memory and inputting the updated embedding vector stored in memory to the decoder, The decoder performs bond prediction and atom prediction, which are predicted as chemical reactions of the plurality of molecular structures, using the updated embedding vector. A program including a command to perform the steps of obtaining the final chemical reaction result predicted from the chemical reaction of the plurality of molecular structures using the bond prediction result and the atom prediction result.

Images