Chemical reaction prediction system and its control method, and learning method of the chemical reaction prediction system

Abstract

The present invention relates to a chemical reaction prediction system and its control method, and a learning method of the chemical reaction prediction system. More particularly, the present invention relates to a chemical reaction prediction system for performing electron flow-based forward reaction prediction and its control method, and a learning method of the chemical reaction prediction system.

Description

Chemical reaction prediction system and control method thereof, and learning method of the chemical reaction prediction system

[0001] The present invention relates to a chemical reaction prediction system, a control method thereof, and a learning method for the 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 the chemical reaction prediction system.

[0002] With the advancement of artificial intelligence, there is a rapid increase in cases where AI technology is achieving outstanding results in various fields.

[0003] In particular, efforts are being made to utilize AI technology to solve various scientific problems, particularly in the natural sciences. For example, in chemistry, active research is being conducted using AI technology to predict the results of intermolecular chemical reactions and design new molecules.

[0004] In this regard, organic synthesis is one of the major challenges in drug development and materials science, and predicting the outcome of chemical reactions is crucial for designing new molecules.

[0005] In the past, a sequence-based model using SMILES strings was used to predict this, but the conversion using SMILES strings has a limitation in that it does not sufficiently reflect the graph structure of the molecule because it is not natural to express the molecular structure in the form of a graph.

[0006] To address these issues, various graph-based approaches have been proposed, representing molecules as graphs. Graph representations can accurately represent molecular structures and provide intuitive information for interpreting and predicting chemical reaction mechanisms. In particular, research utilizing graph neural networks (GNNs) is actively underway, offering the advantage of accurately modeling interactions between atoms within molecules.

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

[0009] The present invention provides a chemical reaction prediction system and a control method thereof, which can accurately predict chemical reactions between molecular structures based on a good understanding of molecular structures, and a learning method of the chemical reaction prediction system.

[0010] In addition, the present invention provides a chemical reaction prediction system and a control method thereof capable of predicting the results of various types of chemical reactions by understanding the chemical reaction mechanism, and a learning method of 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 input and output structures are formed in graph space.

[0012] Furthermore, the present invention provides a learning method of a chemical reaction prediction model that enables a good understanding of chemical reaction mechanisms and generation of more accurate and interpretable chemical reaction prediction results.

[0013] A method for predicting a chemical reaction in which a memory and a processor are cooperatively performed according to the present invention comprises the steps of: receiving information related to a plurality of molecular structures as an input of an encoder; obtaining an embedding vector corresponding to the plurality of molecular structures by using the information related to 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 updated embedding vector through the updating step in a memory, and inputting the updated embedding vector stored in the memory into a decoder; performing bond prediction and atom prediction predicted as a chemical reaction of the plurality of molecular structures by using the updated embedding vector in the decoder; and deriving a final chemical reaction result predicted from a chemical reaction of the plurality of molecular structures by using a result of the bond prediction and a result of the atom prediction. It may include an acquiring step.

[0014] In an embodiment, the step of obtaining the chemical reaction result is characterized by including a step of sampling an initial chemical reaction result using a result of the bond prediction and a result of the atomic prediction, a step of stabilizing the sampled initial chemical reaction result through a diffusion feedback process, and a step of obtaining the final chemical reaction result stabilized through the diffusion feedback process.

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

[0016] In an embodiment, the information related to the plurality of molecular structures includes information about nodes and edges corresponding to the first molecular structure and information about 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 about an atom type (Type), an atomic charge (Charge), a number of hydrogens (Hydrogen), a number of radical electrons (Radical), and a degree (Degree) of a node corresponding to the first molecular structure and a node corresponding to the second molecular structure.

[0017] In an embodiment, in the step of updating the embedding vector, different biases are added to the attention score calculated by the multi-head self-attention layer according to the type of connection between nodes constituting the first molecular graph and the second molecular graph.

[0018] In the embodiment, the bond type is characterized by including a single bond type, a double bond type, a triple bond type, and an aromatic bond type.

[0019] In an embodiment, the method further includes a step of extracting at least one of 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 a shortest path length between nodes constituting the first molecular graph and the second molecular graph, and the K-hop neighbors information includes information on neighboring nodes within K steps for each of the nodes constituting the first molecular graph and the second molecular graph.

[0020] In an embodiment, in the step of updating the embedding vector, the output vector of the multi-head self-attention layer is input to a feed-forward neural network layer, and in the feed-forward neural network layer, the output vector of the multi-head self-attention layer is updated using at least one of the adjacency matrix, the bond type matrix, the shortest paths matrix, and the K-hop neighbors, and the output vector of the feed-forward neural network layer is specified as the updated embedding vector.

[0021] In an embodiment, the combined prediction is performed by performing a dot-product operation using the updated embedding vector, and the dot-product operation is characterized in that it is performed for each vector corresponding to each pair of atoms corresponding to the updated embedding vector.

[0022] In an embodiment, the step of performing the bond prediction includes the step of obtaining an inner product value for each of a plurality of bond types for each of the atom pairs based on the inner product operation, wherein the plurality of bond types are characterized by being associated with at least one of a single bond, a double bond, bond formation, bond collapse, and no change.

[0023] In an embodiment, the step of performing the bond prediction is characterized in that it further includes the step of generating a probability distribution for each of the plurality of bond types for each of the atom pairs using the inner product value according to the inner product operation, and the step of obtaining a transformation matrix that predicts a change in the bonding state of each of the atom pairs using the probability distribution.

[0024] In an embodiment, in the step of generating the probability distribution, a softmax function is applied to the inner product values ​​obtained for the plurality of bonding types for each atom pair, thereby generating the probability distribution for the plurality of bonding types for each atom pair.

[0025] In an embodiment, the step of performing the atomic prediction further includes a step of generating an atomic characteristic probability distribution of each atom corresponding to the updated embedding vector using a softmax output layer, and a step of predicting the atomic characteristics of the atoms corresponding to the updated embedding vector using the probability distribution, wherein the atomic characteristics include charge states of the atoms that may change during a chemical reaction process of the plurality of molecular structures.

[0026] In an embodiment, the diffusion feedback process is characterized by repeatedly evaluating each bond transformation of the initial chemical reaction product and removing or changing unstable bonds.

[0027] In an embodiment, the diffusion feedback process is characterized in that, in order to predict a change in the bonding state of the initial chemical reaction product, the predicted bonding transformation is evaluated for each repeatedly performed step using a transformation probability matrix and a target transformation matrix, and an interpolation factor is used to generate the final chemical reaction product.

[0028] According to the present invention, a method for predicting a chemical reaction in which a memory and a processor are cooperatively performed comprises the steps of: receiving information related to a plurality of 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 plurality of molecular structures, a step of obtaining an embedding vector corresponding to the molecular graph using information related to the plurality of molecular structures in an embedding layer of the encoder, a step of performing an attention operation related to interactions between atoms of the plurality of molecular structures in a multi-head self-attention layer and updating the embedding vector based on the operation, a step of storing the updated embedding vector through the updating step in a memory, and a step of inputting the updated embedding vector stored in the memory to a decoder, a step of performing bond prediction and atom prediction predicted as a chemical reaction of the plurality of molecular structures using the updated embedding vector in the decoder, a step of obtaining a final chemical reaction result predicted from a chemical reaction of the plurality of molecular structures using a result of the bond prediction and a result of the atom prediction, and an actual bond corresponding to the final chemical reaction result and the plurality of molecular structures. The method may include calculating a loss function between label data including states and atomic states, and optimizing a parameter of at least one of the encoder and the decoder to minimize the loss function.

[0029] In a chemical reaction prediction system including a memory, an encoder, a decoder and at least one processor according to the present invention, the encoder receives information related to a plurality of molecular structures, obtains a molecular graph using atoms as nodes and bonds as edges based on the plurality of molecular structures, obtains an embedding vector corresponding to the molecular graph in an embedding layer of the encoder, performs an attention operation related to interactions between atoms of the plurality of molecular structures in a multi-head self-attention layer of the encoder, and updates the embedding vector based on the operation, the processor stores the updated embedding vector through the updating step in a memory, and inputs the updated embedding vector stored in the memory to a decoder, and the decoder performs bond prediction and atom prediction predicted as a chemical reaction of the plurality of molecular structures using the updated embedding vector, and compares the result of the bond prediction and the result of the atom prediction. Using the results, the final chemical reaction product predicted from the chemical reaction of the above multiple molecular structures can be obtained.

[0030] A program stored in a computer-readable medium, which is executed by one or more processes in an electronic device according to the present invention, wherein the program comprises: a step of receiving information related to a plurality of molecular structures as an input of an encoder; a step of obtaining a molecular graph using atoms as nodes and bonds as edges based on the plurality of molecular structures; a step of obtaining an embedding vector corresponding to the molecular graph in an embedding layer of the encoder; a step of performing an attention operation related to interactions between atoms of the plurality of molecular structures in a multi-head self-attention layer and updating the embedding vector based on the operation; a step of storing the embedding vector updated through the updating step in a memory, and inputting the updated embedding vector stored in the memory into a decoder; a step of performing bond prediction and atom prediction predicted as a chemical reaction of the plurality of molecular structures using the updated embedding vector in the decoder; and a step of performing the bond prediction and atom prediction. It may include instructions for performing a step of obtaining a final chemical reaction result predicted from a chemical reaction of the plurality of molecular structures using the result of the prediction and the result of the atomic prediction.

[0031] As described above, according to the chemical reaction prediction system and control method thereof according to the present invention, and the learning method of the chemical reaction prediction system, the movement of electrons is modeled through the graph structure of molecules, thereby enabling a good understanding of the chemical reaction mechanism and accurate prediction of the chemical reaction result.

[0032] Furthermore, the chemical reaction prediction system and its control method according to the present invention, as well as the learning method for the chemical reaction prediction system, provide a chemical reaction prediction model that enables a thorough understanding of chemical reaction mechanisms and accurate prediction of chemical reaction results, thereby reducing the time and cost required for experiments by users. This reduces research and development costs and shortens the time to market for new products.

[0033] Meanwhile, according to the chemical reaction prediction system and control method thereof according to the present invention, and the learning method of the chemical reaction prediction system, both the input of the molecular structure and the output of the predicted result are performed in the graph space, thereby enabling forward reaction prediction to be performed regardless of the SMILES permutation and order.

[0034] Furthermore, according to the chemical reaction prediction system and control method thereof according to the present invention, and the learning method of the chemical reaction prediction system, it is possible to simultaneously sample multiple interdependent transformations occurring in parallel within a molecular graph, thereby ensuring consistency between transformations.

[0035] Furthermore, according to the chemical reaction prediction system and control method thereof according to the present invention, and the learning method of the chemical reaction prediction system, in order to solve problems that may arise due to a symmetrical structure, symmetry is broken by including a noise or sampling mechanism, and a valid output structure is formed, thereby preventing the occurrence of an invalid configuration.

[0036] FIG. 1 is a conceptual diagram for explaining a chemical reaction prediction system and a control method thereof according to the present invention, and an answer generation system to which a learning method of the chemical reaction prediction system is applied.

[0037] FIG. 2a, FIG. 2b, FIG. 2c, FIG. 2d, FIG. 2e and FIG. 2f are conceptual diagrams for explaining a chemical reaction prediction model according to the present invention.

[0038] Figure 3 is a flowchart for explaining a learning method of a chemical reaction prediction model according to the present invention.

[0039] Figures 4a, 4b, 4c and 4d are conceptual diagrams for explaining a chemical reaction prediction model according to the present invention.

[0040] FIG. 5a, FIG. 5b, FIG. 5c, FIG. 5d, FIG. 5e, FIG. 6 and FIG. 7 are conceptual diagrams for explaining examples of use in an answer system to which a chemical reaction prediction model according to the present invention is applied.

[0041] Hereinafter, embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Regardless of the drawing numbers, identical or similar components will be given the same reference numbers, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" used for components in the following description are assigned or used interchangeably only for the convenience of writing the specification, and do not in themselves have distinct meanings or roles. In addition, when describing the embodiments disclosed in this specification, if it is determined that a specific description of a related known technology may obscure the gist of the embodiments disclosed in this specification, a detailed description thereof will be omitted. In addition, the attached drawings are only intended to facilitate easy understanding of the embodiments disclosed in this specification, and the technical ideas disclosed in this specification are not limited by the attached drawings, and should be understood to include all modifications, equivalents, and substitutes included in the spirit and technical scope of the present invention.

[0042] Terms that include ordinal numbers, such as first, second, etc., may be used to describe various components, but the components are not limited by these terms. These terms are used solely to distinguish one component from another.

[0043] When a component is referred to as being "connected" or "connected" to another component, it should be understood that it may be directly connected or connected to that other component, but that there may be other components intervening. Conversely, when a component is referred to as being "directly connected" or "connected" to another component, it should be understood that there are no other components intervening.

[0044] Singular expressions include plural expressions unless the context clearly indicates otherwise.

[0045] In this application, terms such as “include” or “have” are intended to specify the presence of a feature, number, step, operation, component, part or combination thereof described in the specification, but should be understood not to exclude in advance the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts or combinations thereof.

[0046] Hereinafter, the present invention will be described in more detail with reference to the attached drawings. Fig. 1 is a conceptual diagram for explaining a chemical reaction prediction system according to the present invention, a control method thereof, and an answer generation system to which a learning method of the chemical reaction prediction system is applied. Figs. 2a, 2b, 2c, 2d, 2e, and 2f are conceptual diagrams for explaining a chemical reaction prediction model according to the present invention, and Fig. 3 is a flowchart for explaining a learning method of the chemical reaction prediction model according to the present invention. Furthermore, Figs. 4a, 4b, 4c, and 4d are conceptual diagrams for explaining a chemical reaction prediction model according to the present invention, and Figs. 5a, 5b, 5c, 5d, 5e, 6, and 7 are conceptual diagrams for explaining an example of use in an answer system to which a chemical reaction prediction model according to the present invention is applied.

[0047] The chemical reaction prediction system, its control method, and its learning method according to the present invention can be usefully utilized in various situations. Predicting chemical reactions is utilized in various ways, and for example, it can be usefully utilized in research for designing new materials or developing new drugs. In this regard, organic synthesis is one of the important tasks in new drug development and materials science, and predicting the results of chemical reactions is very important in designing new molecules. The present invention provides a system, a control method, and a learning method for predicting chemical reactions between molecular structures by transforming molecular structures into graphs and predicting chemical reactions based on the graphs. The chemical reaction prediction system, its control method, and its learning method according to the present invention are implemented based on a “chemical reaction prediction model.” For convenience of explanation, the system, method, and learning method are not separately named, but are collectively referred to as a “chemical reaction prediction” model.

[0048] Meanwhile, the chemical reaction prediction model according to the present invention can be applied to various industries and services, and for example, it can be usefully utilized in answer generation systems based on language models. Recently, along with the advancement of deep learning technology, generative AI technology has been attracting attention. More specifically, generative AI models can generate new data in various forms, such as text, images, and voice, from given data, providing a different level of application potential than simply classifying or predicting existing data. The chemical reaction prediction model according to the present invention can also be applied to such answer generation systems and be usefully utilized in various fields that require chemical reaction prediction, such as research for designing new materials or developing new drugs.

[0049] Hereinafter, a response generation system to which a chemical reaction prediction model according to the present invention can be applied will be briefly described with reference to FIG. 1. The response generation system according to FIG. 1 may include various prediction and analysis models, and may be a system that uses these to generate property prediction results for molecular structures or to design molecules having characteristics desired by a user. In addition, the response generation system according to the present invention may be a system that generates chemical reaction prediction results for new types of molecules and / or between multiple molecules. Furthermore, the response generation system according to the present invention may be a system that generates prediction results for modifications of existing materials and synthesis of various materials (e.g., new materials, polymer materials, nanomaterials, composite materials, organic materials, pharmaceutical materials, etc.).

[0050] The answer generation system according to the present invention includes a large-scale foundation model (or a large-scale foundation artificial intelligence model, or a generative artificial intelligence model), and the present invention aims to increase the efficiency of natural science research by minimizing the risk of research failure. This answer generation system may also be referred to as an answer generation platform based on a large-scale foundation model. However, the "large-scale foundation model" may also be referred to as a generative model, a foundation model, or a large language model (LLM).

[0051] Meanwhile, as illustrated in FIG. 1, the answer generation system (100) may include at least one of an input unit (110), an output unit (120), a communication unit (130), a storage unit (140), and a super-large foundation model (200). Here, the super-large foundation model (200) may also be referred to as a foundation model, and the foundation model may mean a super-large AI core base model trained with a massive dataset.

[0052] Although not shown, the answer generation system (100) may include one or more processors, which 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 integrated circuits (ASICs), etc.). The one or more processors may be configured to execute instructions, computer-readable instructions, and / or other instructions described herein stored (or included) in the storage (140). The answer generation system and method may enable the memory and at least one processor to cooperate to perform the data processing described below. The processor may perform a series of operations and data processing using data and information stored in the memory. In this case, the memory may be a component of the storage (140).

[0053] Meanwhile, the input unit (110) may be configured as a means for data input and may be configured in various types. 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 a user terminal (10). Here, “receiving input” may mean receiving an input signal (or selection signal) corresponding to the user’s input based on the input being made by the user through the input unit configuration provided in the user terminal (10).

[0054] The input unit (110) may also be referred to as a user interface module. The input unit (110) may include a touch screen, a computer mouse, a keyboard, a keypad, a touch pad, a trackball, a joystick, a voice recognition module, or other similar devices. However, the present invention does not place any limitations on the type of input unit (110). Furthermore, the input unit (110) in the present invention does not necessarily refer to a hardware means, but can be understood as a channel for receiving input from a user.

[0055] Here, the user input may include documents, text, images (or videos), voice, etc. In this case, the answer generation system (100) may further include a module that converts voice into text.

[0056] Next, the output unit (120) can output information through an output unit configuration (e.g., a display unit, a touch screen, a speaker, etc.) provided in a user terminal (10) linked to the answer generation system (100). For example, the output unit (120) can output a page (or service page, 1000) linked to the answer generation system (100) to the display unit of the user terminal (10). In addition, the output unit (120) does not necessarily mean a hardware means, and can be understood as a channel for outputting results to the user.

[0057] Next, the communication unit (130) may be connected to a user terminal (10), a server (e.g., a central server, an external server, etc.), a device, and at least one network through a wireless or wired network, and may be configured to receive or transmit overall data and information necessary for the operation of the answer generation system (100).

[0058] Here, the user terminal (10) may include at least one of a mobile phone, a smart phone, a notebook computer, a laptop computer, a slate PC, a tablet PC, an ultrabook, a desktop computer, a digital broadcasting terminal, a personal digital assistant (PDA), a portable multimedia player (PMP), a navigation device, and a wearable device (e.g., a smartwatch, a smart glass, a head mounted display (HMD)).

[0059] Furthermore, the communication unit (130) can support various communication methods according to the communication standards of the communicating device.

[0060] For example, the communication unit (130) may be configured to communicate with a communication target using at least one of 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), Bluetooth™, RFID (Radio Frequency Identification), Infrared Data Association (IrDA), UWB (Ultra-Wideband), ZigBee, NFC (Near Field Communication), Wi-Fi Direct, and Wireless USB (Wireless Universal Serial Bus) technologies.

[0061] Meanwhile, the storage unit (140) serves to store various data related to the present invention, and may include one or more non-transitory computer-readable storage media that can be read and / or accessed by at least one of the one or more processors (140).

[0062] The 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 (140) may be implemented using a single physical device (e.g., a single optical, magnetic, organic, or other memory or disk storage device), while in other examples, the storage (140) may be implemented using two or more physical devices.

[0063] 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 devices and networks.

[0064] Furthermore, at least a portion of the storage unit (140) may be a cloud storage or a cloud server. At least a portion of data corresponding to user input received from the input unit (110) and learning data may be stored in the storage unit (140).

[0065] That is, it can be understood that the storage unit (140) is sufficient as a space in which information necessary for the operation of the answer generation system (100) is stored, and there are no restrictions on the physical space.

[0066] Meanwhile, the super-large foundation model (200) can be configured to generate property prediction results for molecular structures or to design molecules with desired characteristics by the user. Furthermore, the super-large foundation model (200) can generate prediction results for new types of molecules and / or chemical reactions between multiple molecules, or to generate prediction results for the modification of existing materials and the synthesis of various materials (e.g., new materials, polymer materials, nanomaterials, composite materials, organic materials, pharmaceutical materials, etc.).

[0067] In this regard, the super-large foundation model (200) may include at least one of a document understanding model (300), a chemical reaction prediction model (400), and a molecular property prediction model (500).

[0068] A document understanding model (300) can extract various types of content that satisfy preset content criteria from documents (e.g., papers, books, patent documents, reports, etc.). More specifically, the document understanding model (300) can 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 extract various contents (or data) and knowledge based on the understood contents.

[0069] Here, the preset content criteria can be set in various ways and can be determined according to the purpose or use of the answer generation system (100) according to the present invention.

[0070] For example, if the purpose of utilizing the answer generation system (100) is chemistry, bio, new materials, new substances, and new drug development, the document understanding model (300) can be trained to understand and extract content related to chemistry, bio, new materials, new substances, and new drug development from the document to be analyzed.

[0071] In this case, the preset content criteria may include content related to molecular structures related to at least one of chemistry, biology, new materials, new substances, and new drug development. Here, the document understanding model (300) can extract content related to chemistry, biology, new materials, new substances, and new drug development from the analysis target document based on the preset content criteria.

[0072] For convenience of explanation, in this specification, the established content criteria are explained as being related to at least one of chemistry, bio, new materials, new substances, and new drug development, but are not necessarily limited thereto.

[0073] The document understanding model (300) can extract at least one of text, molecular structure, formula, chart, table, and image that satisfies the preset content criteria from the document to be analyzed based on the preset content criteria.

[0074] For example, the document understanding model (300) can understand the chemical structure of the molecular structure included in the analysis target document (20), and based on the result of the understanding, can convert the molecular structure into a SMILES string expression and extract it. In addition, the document understanding model (300) can understand the chemical structure of the molecular structure, and based on the result of the understanding, can perform a graph transformation corresponding to the molecular structure.

[0075] Meanwhile, in the present invention, receiving "molecular structure" can be understood as receiving information that can identify a molecule. In this case, the information that can identify a molecular structure can take various forms, such as a molecular structural formula, a molecular graph, a chemical formula, a molecular structural formula according to the SMILES notation, or a molecular structure image.

[0076] In addition, the document understanding model (300) can understand text (23) related to the molecular structure (21) among the texts included in the document to be analyzed (20) and extract it as text data (23) related to the molecular structure (21).

[0077] Furthermore, the document understanding model (300) can recognize rows and columns forming a table (24) associated with a molecular structure (21) from the document to be analyzed (20), and extract them by converting them into structured data (25) in a format such as HTML or Excel.

[0078] Furthermore, the document understanding model (300) can extract relationship information (or relationships) between molecular structures included in the document.

[0079] As discussed above, the document understanding model (300) can extract various types of data contained in a document by converting them into data in a form (e.g., machine-readable data) that the super-large foundation model (200) can understand. The data extracted using the document understanding model (300) can be sorted by page or document unit and stored in a storage unit (or memory, 140). In the present invention, the document understanding model (300) can also be referred to as a deep document understanding model.

[0080] Next, the chemical reaction prediction model (400) can be configured to predict a chemical reaction between input molecular structures (or compounds or reactants) and output (or generate) a predicted chemical reaction result (or product). For example, as illustrated in FIG. 2A, when information on the molecular structure (401a, 401b) of a specific compound is input, the chemical reaction prediction model (400) can predict a product (402) formed by a chemical reaction between the molecular structure (401a, 401b) of the specific compound.

[0081] The chemical reaction prediction model (400) according to the present invention can model the movement of electrons through the graph structure of molecules. The chemical reaction prediction model (400) according to the present invention can be an electron flow-based prediction model using graph diffusion that can well understand the chemical reaction mechanism and accurately predict the results of a chemical reaction. The chemical reaction prediction model (400) can express both inputs and outputs in a graph space. The chemical reaction prediction model (400) can output a reaction product and / or byproduct prediction result based on the graph structure of an input compound by learning the graph transformation of a chemical reaction. The chemical reaction prediction model (400) in the present invention can also be called an electron flow model, an electron flow-based forward reaction prediction model, or an electron-flow inspired graph diffusion model for interpretable forward reaction prediction using a graph diffusion model that simulates electron flow.

[0082] Meanwhile, the chemical reaction prediction model (400) discussed above may include at least one of an encoder (410) and a decoder (420).

[0083] The encoder (410) according to the present invention can be configured to extract the characteristics of each atom and bond through a graph representation of a molecular structure and generate an embedding vector based on the extracted characteristics. In addition, by performing an attention operation related to the interaction between atoms of the plurality of molecular structures through a neural network, the characteristics of the molecular structure can be better understood. 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 vector obtained from the encoder (410). The chemical reaction prediction model (400) can perform a diffusion feedback process to stabilize the chemical reaction results according to the bond prediction and atom prediction, and generate the final chemical reaction results. The configuration and detailed process of the chemical reaction prediction model (400) according to the present invention will be described in more detail later.

[0084] Meanwhile, the chemical reaction prediction model (400) according to the present invention can utilize not only molecular structural information but also various chemical information such as reaction conditions, catalysts, and temperature to more accurately predict chemical reactions. This chemical information may include molecular structural data (e.g., molecular structure graph) and descriptive data (e.g., text data). Since chemical reactions are affected by various factors such as molecular structural changes, reaction mechanisms, reaction conditions, and reaction environments, more accurate predictions can be possible only when these factors are comprehensively considered. Meanwhile, such chemical information may be included in various literature, textbooks, theses, patent literature, articles, and academic journals, and the chemical reaction prediction model (400) according to the present invention can perform more accurate chemical reaction predictions by integrating such chemical information.

[0085] To this end, as illustrated in FIG. 2b, the chemical reaction prediction model (400) may include at least one of a first module (400a) and a second module (400b). Here, the first module (400a) may also be named a “ChemExpert-Graph” module, a graph module, a graph processing module, a graph model, a graph processing model, etc., and the second module (400b) may also be named a “ChemExpert-Text” module, a text module, a text processing module, a text model, a text processing model, etc.

[0086] The first module (400a) can receive a molecular (or chemical) structure as input and predict a graph-based chemical reaction. For example, referring to (a) of FIG. 2d, an example of a chemical reaction can be seen. Furthermore, as illustrated in (b) of FIG. 2d, the first module (400a) can include multiple layers (400a-1) to predict a graph-based chemical reaction. More specific details regarding the multiple layers (400a-1) will be described later.

[0087] The molecular structure (403a, or molecular structural formula) input to the first module (400a) is converted into a molecular graph in the form of a graph, and in the molecular graph, atoms can be expressed as nodes and bonds can be expressed as edges.

[0088] The molecular structure (403a) may correspond to at least one of data extracted from a document (403) including a molecular structure (411) using a document understanding model (300) or information extracted through a storage unit (or memory, 140).

[0089] The first module (400a) can analyze changes in the structural properties of a molecule based on an input molecular graph, predict a chemical reaction path and a product to be generated as a result of the chemical reaction, and output the predicted chemical reaction path and product.

[0090] In one embodiment, the first module (400a) can analyze structural changes of a molecule based on a molecular graph to predict the process in which a specific bond is broken and a new bond is formed.

[0091] In another embodiment, the first module (400a) can analyze the interactions between atoms within a molecule based on a molecular graph and predict radical formation and bonding changes that may occur at each step.

[0092] That is, the first module (400a) can be configured to receive a molecular graph as input and output a chemical reaction path and product (404a) predicted based on the molecular graph.

[0093] Next, the second module (400b) may be configured to process text data (403b) to understand and predict the reaction mechanism. In this case, the text data (403b) may be data extracted from a document containing a molecular structure (403a) using a document understanding model (300), or may correspond to at least one piece of information extracted through a storage unit (or memory, 140) related to the molecular structure (411). The second module (400b) may be a model that has learned a large amount of data related to chemical reactions.

[0094] In one embodiment, the text data (403b) input to the second module (400b) is data including a description of the molecular structure (403a), and may include at least one of chemical reaction conditions, chemical reaction mechanism (or reaction path), and chemical properties of the molecular structure (403a).

[0095] 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 chemical reaction conditions, chemical reaction mechanism, and chemical properties predicted based on the text data (403b).

[0096] In one embodiment, the second module (400b) may analyze text data (403b) using natural language processing (NLP) technology and extract at least one text among chemical reaction conditions, chemical reaction mechanism (or reaction path), chemical properties, and experimental data included in the text data (403b).

[0097] In another embodiment, the second module (400b) can predict a chemical reaction mechanism (e.g., how a specific catalyst or condition affects a reaction) based on text extracted through analysis of text data (403b), and output the predicted chemical reaction mechanism and chemical properties.

[0098] The second module (400b) can analyze information related to prediction of chemical reactions related to the plurality of molecular structures from text data.

[0099] The chemical reaction prediction model (400) can combine the output data (404a) of the first module (400a) and the output data (404b) of the second module (400b) to output the final chemical reaction prediction result (e.g., product, chemical reaction path, chemical reaction mechanism, etc.).

[0100] In one embodiment, as illustrated in (a) to (c) of FIG. 2c, 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 from the first module (400a) and the second module (400b). In this case, the electron flow, reaction conditions, and structural effects can be displayed together as graphs and text, and a molecular structure reflecting the positions before and after electron movement can be generated, or the molecular structure of a product generated according to the reaction conditions can be generated.

[0101] That is, the chemical reaction prediction model (400) can make more accurate predictions than using only a single data source by fusing and utilizing output data (404a, 404b) output from each of the first module (400a) and the second module (400b), and can enable users to intuitively recognize various elements related to chemical reactions.

[0102] In addition, the chemical reaction prediction model (400) can verify the chemical reaction result predicted by the first module (400a) using the output data analyzed by the second module (400b). That is, the second module (400b) can obtain at least one of the chemical reaction conditions, chemical reaction mechanism, and chemical characteristics analyzed based on the text data (403b). The chemical reaction prediction model (400) can verify whether the chemical reaction result predicted and obtained by the first module (400a) matches experimental data or theoretical expectations based on the data analyzed by the second module (400b).

[0103] Meanwhile, the chemical reaction prediction model (400) according to the present invention is characterized by performing chemical transformation based on a molecular graph, and for understanding, the composition of the molecular structure graph will be briefly examined.

[0104] Figure 2e (a) represents nodes and edges, and in a molecular graph, atoms can be represented as nodes and bonds between atoms as edges. For example, as shown in Figure 2e (b), a water molecule has a molecular formula of “H2O”, and at this time, the atoms are composed of H, H, and O. In this case, there are three nodes, and as shown in Figure 2e (c), the atoms H, H, and O can be represented by modes n1, n2, and n3, respectively. In addition, there can be two bonds between atoms, OH and OH, and these bonds can be represented by edges (e1, e2), as shown in Figure 2e (c). In this way, when a molecular structure is converted into a molecular graph, the unique positional and topological relationships of the molecular structure can be preserved, enabling more accurate prediction. In the present invention, a molecular structure to be predicted is converted into a molecular graph and input into an encoder (410) to predict a chemical reaction.

[0105] Meanwhile, in the present invention, the encoder (410) can embed a molecular structure into a vector using a molecular graph. At this time, briefly looking at the vector, as illustrated in (a) of Fig. 2f, an ethanol (C2H5OH) molecule can represent ethanol through a vector having a specific dimension, as illustrated in (b) of Fig. 2f, and for example, a vector corresponding to each atom of ethanol can be expressed as shown in the figure. At this time, the dimension of the vector and the information included can be set in various ways. For example, the figure represents a five-dimensional vector, and each vector can include atom type, bonding information, charge information, hybridization information, directionality information, etc. Meanwhile, although not illustrated, the encoder (410) can embed a vector for the bonding information of the molecule and utilize it for analysis.

[0106] Below, a method for predicting chemical reactions and obtaining chemical reaction results based on molecular graphs using a chemical reaction prediction model (400) according to the present invention will be described in more detail. The description below can be performed in the first module (410) illustrated in FIG. 2b.

[0107] Chemical reaction prediction and organic synthesis are key challenges in drug discovery and / or materials science. Predicting the outcome of chemical reactions is crucial for designing novel molecules and can significantly contribute to shortening product development cycles in various industries. 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) as ASCII strings, allowing them to be processed as text data using natural language processing techniques.

[0108] However, SMILES strings do not match the natural graph representation of molecules, which reduces learning efficiency and can generate chemically invalid transformations (or transformations). Specifically, for chemical reactions, transformations in SMILES space may not directly correspond to valid molecular graphs, resulting in chemically infeasible outputs. In other words, these transformations have the limitation that they do not always lead to valid transformations in graph space, a more natural representation of molecules.

[0109] To address these issues, various graph-based (or graph-centric) approaches have been proposed to represent molecules as graphs. Graphical representations can accurately represent molecular structures, enhance the interpretability of chemical reaction mechanisms, and facilitate more accurate predictions of chemical reactions based on electron flow theory.

[0110] Accordingly, the present invention proposes a chemical reaction prediction model based on electron flow using graph diffusion, which models the movement of electrons through the graph structure of molecules, thereby enabling a thorough understanding of chemical reaction mechanisms and accurate prediction of chemical reaction results. The prediction method according to the present invention may also be called an interpretable forward reaction prediction method using a graph diffusion model that simulates electron flow (Electron-Flow Inspired Graph Diffusion Model for Interpretable Forward Reaction Prediction).

[0111] The chemical reaction prediction model described in the present invention utilizes graph diffusion, with both input and output operating in graph space. This maintains invariance with respect to the permutation and order of SMILES strings, while providing users with more interpretable transformations. Rather than starting with random graph sampling, the chemical reaction prediction model according to the present invention can learn the transformation process from an initial set of graphs to the final product.

[0112] Below, the present invention will be examined in more detail with reference to the attached drawings.

[0113] As illustrated in FIG. 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 illustrated, the chemical reaction prediction model (400) may perform a series of data processing according to the present invention in cooperation with a memory and at least one processor. Various information may be stored in the memory. For example, the memory may store various information such as information input to the chemical reaction prediction model (400), results generated according to a series of data processing in the chemical reaction prediction model (400), information on intermediate products, and information on final results. The processor may cooperate with each component of the chemical reaction prediction model to allow the process for chemical reaction prediction according to the present invention to proceed in each component and neural network layer.

[0114] In the chemical reaction prediction model (400) according to the present invention, during the learning stage, as shown in FIGS. 4a and 4b, a learning unit (440) and a ground-truth product exist, but after learning is completed, as shown in FIGS. 4c and 4d, the learning unit and the ground-truth product may be excluded.

[0115] 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). Meanwhile, the diffusion process described in the present invention does not depend on a backbone network and can be implemented with various encoders.

[0116] The encoder (410) according to the present invention may include an embedding layer (411) and a graph neural network (GNN, Graph Neural Networks, 412), as illustrated. The graph neural network may be an attention-based graph neural network or a backbone neural network.

[0117] Meanwhile, 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 may also be referred to as a topologically weighted multi-head self-attention layer.

[0118] The embedding layer (411) is configured to receive information related to a plurality of molecular structures and convert it into an embedding vector. The embedding layer (411) may be configured to generate a vector including atoms of the molecular structure and bonding relationships between atoms. The graph neural network (412) may receive the embedding vector embedded in the embedding layer (411), perform an attention operation related to the interaction between atoms of the plurality of molecular structures, and update the embedding vector based on the operation. At this time, the embedding layer may receive a molecular graph and obtain an embedding vector based on the molecular graph.

[0119] Meanwhile, node features contained in the embedding vector output by the embedding layer (411) may include atom type, atomic charge, the number of implicit hydrogens, and the number of radical electrons. The embedding layer (411) can encode nodes according to the molecular graph to generate an embedding vector containing information usable within the encoder. This allows the molecular graph to be fully reconstructed into individual molecules, and the node degrees can also be tokenized as input features and later used to measure the consistency between the bond prediction and the predicted atomic degrees.

[0120] The features of the atom v can be represented as (f), (g), and (h) of Fig. 5a, and can be tokenized and embedded through the embedding layer (411).

[0121] As illustrated in (g) of FIG. 5a, the features of the atom v are tokenized and embedded through the embedding layer (411), and the tokenized features of type i according to (f) of FIG. 5a may have features such as those illustrated in (g) of FIG. 5a. i is an index indicating the type of the atomic feature, and may belong to the set of type, charge, hydrogen, radical, and order. i may mean at least one of the features of type, charge, hydrogen, radical, and order, where type may mean the atom 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 degree of the node (the number of other atoms bonded to the atom). These features are embedded through the embedding layer (411), as illustrated in (h) of FIG. 5a, to form an initial embedding of the atom. The initial embedding of atom v can be the average of all embeddings.

[0122] The embedding layer (411) can perform embedding for multiple molecules based on a molecular graph. The molecular graph for the molecular structure may be stored in memory, a dataset, a database, etc. Alternatively, the molecular graph for the 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) or may cooperate with the molecular graph extraction module (450). Furthermore, the molecular graph extraction module (450) may of course be included as a component of the encoder (410).

[0123] A molecular graph can maintain multiple molecular structures by representing atoms as nodes and bonds between atoms as edges.

[0124] The structure of the plurality of molecules may include a first molecular structure and a second molecular structure, and these first and second molecular structures may be referred to as reactants. They correspond to the drawing symbols 401a and 401b in Fig. 4a.

[0125] The molecular graph extraction module (450) can obtain a molecular graph for the plurality of molecular structures by converting atoms into nodes and bonds between atoms into edges based on the plurality of molecular structures using a pre-specified graph transformation algorithm.

[0126] The molecular graph extraction module (450) can obtain 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 bonding relationships between 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 atoms constituting the second molecular structure into nodes and bonding relationships between atoms constituting the second molecular structure into edges using the specified graph transformation algorithm.

[0127] In the embedding layer (411) of the encoder (410), an embedding vector corresponding to the plurality of molecular structures can be obtained using information related to the plurality of molecular structures. At this time, the information related to the plurality of molecular structures can include information on nodes and edges corresponding to the first molecular structure and information on nodes and edges corresponding to the second molecular structure. The information related to the molecular structure can be a molecular graph extracted by the molecular graph extraction module (450) or information extracted from such a molecular graph.

[0128] The embedding vector may include at least one piece of information about the atomic type (Type), atomic charge (Charge), number of hydrogens (Hydrogen), number of radical electrons (Radical), and degree (Degree) of the node corresponding to the first molecular structure and the node corresponding to the second molecular structure. Furthermore, the embedding vector may be configured to further include information about the bonding relationship between atoms of the first and second molecular structures.

[0129] Meanwhile, when the reactants to be predicted, i.e., a plurality of molecular structures (e.g., the first molecular structure and the second molecular structure, 401a, 401b) are specified, the chemical reaction prediction model (400) can extract at least one of an adjacency matrix, a bond type matrix, a shortest path, and K-hop neighbors corresponding to the first molecular graph and the second molecular graph, respectively, by using the nodes and edges corresponding to the first molecular graph and the nodes and edges corresponding to the second molecular graph. The time point at which such information (480) is extracted is not limited and may vary.

[0130] The above adjacency matrix may include information about direct connections between nodes constituting the first molecular graph and the second molecular graph.

[0131] The above coupling type matrix may include information on coupling types between nodes constituting the first molecular graph and the second molecular graph.

[0132] The above shortest path matrix may include information about the shortest path length between nodes constituting the first molecular graph and the second molecular graph.

[0133] The above K-step neighbor information may include information about neighboring nodes within K steps for each of the nodes constituting the first molecular graph and the second molecular graph.

[0134] At least one (480) of the adjacency matrix, the bond type matrix, the shortest paths, and the K-hop neighbors extracted in this manner can be stored in memory. At least one of the adjacency matrix, the bond type matrix, the shortest paths, and the K-hop neighbors stored in the memory can be input to at least one of the embedding layer (411), the neural network (412), the multi-head self-attention layer (413), and the feedforward neural network (414).

[0135] In the embedding layer (411), when an embedding vector (v1) corresponding to a plurality of molecular structures (or reactants, for example, the first molecular structure, the second molecular structure, 401a, 401b) is obtained, the obtained embedding vector (v1) can be input to a graph neural network (412). In particular, the embedding vector (v1) can be input to a multi-head self-attention layer (413) of the graph neural network.

[0136] The multi-head self-attention layer (413) can perform attention operations related to 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 operations. At this time, the embedding vector (v1) to be updated may be a vector output from the embedding layer.

[0137] The multi-head self-attention layer (413) can perform global attention-based pooling. To account for the molecular graph topology, the multi-head self-attention layer (413) can add a bias to the attention weight scores derived from the shortest path between pairs of nodes. This can act as an encoding of the relative positions between atoms. Consequently, this can be free from spectral-based encoding issues (e.g., lack of sign invariance for Laplacian matrix eigenvectors) because the distance measurement between nodes is not absolute. This encoding is computationally efficient, allowing global message passing while considering the entire graph topology. At this time, the calculation time of the shortest path can be varied.

[0138] Global attention-based pooling is a method that considers the entire topology of a graph, and the multi-head self-attention layer (413) can globally consider the correlation between each node of the graph by performing global attention-based pooling.

[0139] For example, assuming that the dimension of the input vector from the embedding layer is 128 and the number of heads in the multi-head attention is 8, each head can be input with a 16-dimensional sub-vector, which is obtained by dividing the input vector by the 128 dimension, the number of heads, 8.

[0140] Each head of the multi-head self-attention layer (413) is configured to independently calculate an attention value for a 16-dimensional sub-vector by applying a preset attention mechanism. For example, the first head may mainly reflect information of adjacent nodes in the molecular graph, and the second head may mainly reflect information of a specific combination type in the molecular graph. The 16-dimensional attention vectors calculated from multiple heads, for example, eight heads, of the multi-head self-attention layer (413) may be combined again to form a 128-dimensional vector. This combined vector has the same dimension as the embedding vector output from the embedding layer (411), but may include richer information by integrating various information extracted from each head. In the present invention, this may be expressed as an updated embedding vector.

[0141] Meanwhile, in order to reflect the shortest path between nodes in the weights, in the present invention, as discussed above, the shortest path between all pairs of nodes in the molecular graph can be calculated. The shortest path can be calculated using a preset algorithm. The multi-head self-attention layer (413) can use the calculated shortest path information to encode the relative positions between nodes. This encoding can be performed based on the shortest path distance between nodes. The multi-head self-attention layer (413) can use the shortest path information as a bias when calculating the attention weight score between nodes. For example, the multi-head self-attention layer (413) can adjust the attention weight by increasing the attention weight as the shortest path distance between two nodes is close, and decreasing it as the shortest path between two nodes is farther away. The multi-head self-attention layer (413) can add the bias derived from the shortest path between pairs of nodes to the attention weight score.

[0142] The multi-head self-attention layer (413) encodes each node of the molecular graph using attention weights adjusted based on the shortest path, thereby enabling encoding that takes into account the topology of the entire molecular graph.

[0143] Additionally, in the multi-head self-attention layer (413), a bias can be added to the attention score depending on the combination type (e.g., single, double, triple, aromatic combination).

[0144] For example, in the multi-head self-attention layer (413), different biases can be added to the attention score calculated by the multi-head self-attention layer according to the type of connection between the nodes constituting the first molecular graph and the second molecular graph.

[0145] Here, the bond type may include a single bond type, a double bond type, a triple bond type, and an aromatic bond type.

[0146] This may mean that instead of a single bias matrix for all heads, each head has its own bias matrix. For example, the first head may have a shortest-path bias, and the second head may have a bias based on a single bond. The remaining heads may use masks that focus on global attention between atoms or molecules. In the multi-head self-attention layer (413) of the present invention, these various bias matrices may be added to each layer corresponding to the multi-head. In the present invention, by adding a bias matrix to each layer corresponding to the multi-head, the attention score of each head may emphasize information masked by the bias matrix. For example, when a bias matrix based on a bond type is added, the head is masked based on the bond type and can only receive information from neighbors connected by a specific bond type. For example, a specific head may receive information from nodes connected only by a single bond.

[0147] In this way, the weights for the attention score according to the present invention, which add biases reflecting various characteristics of the molecular structure, can be calculated according to the formula and meaning shown in Fig. 5. Here, A(Qk,Kk,Vk) can mean the attention weights calculated by applying a softmax function to the attention score.

[0148] In this way, in the multi-head self-attention layer (413), when an attention operation related to the interaction between atoms of the plurality of molecular structures is performed and the embedding vector is updated based on the operation, the updated embedding vector can be stored in memory through the update step. Then, the updated embedding vector stored in the memory can be input to a feed-forward neural network layer (414). The feed-forward neural network layer (414) can receive the output of the multi-head attention layer (413) and additionally 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, the bond type matrix, the shortest paths, and the K-hop neighbors. The above feedforward neural network layer (414) can update the output vector of the multi-head self-attention layer to include more structural information about molecular structures. The output vector of the feedforward neural network layer can be specified as the final updated embedding vector (v2) input to the decoder (420).

[0149] Furthermore, the decoder (420) according to the present invention may include at least one projection layer (Projection Layer, 421). In the decoder (420), bond prediction and atom prediction based on a plurality of molecular structures (reactants) may be performed based on the updated embedding vector (v2) input to the projection layer. Bond prediction is to predict bonds between atoms, and atom prediction is to predict state changes of atoms. The decoder (420) may include a plurality of projection layers (422a, 422b). In each projection layer (422a, 422b), bond prediction (Bond Prediction, 423) and atom prediction (Atom Prediction, 424) may be performed.

[0150] In the decoder (420), joint prediction can be achieved by performing a dot product operation using the updated embedding vector. The dot product operation can be performed for each vector corresponding to each pair of atoms corresponding to the updated embedding vector.

[0151] In block 423 of the drawing, each box on the x-axis and y-axis (423a, 423b) conceptually represents each atom of the molecules to be reacted, and these atoms can be represented by the updated embedding vector (v2). The decoder (420) can predict the bonding type between each atom through an inner product operation based on the information contained in the updated embedding vector.

[0152] The decoder (420) can obtain inner product values ​​for each of multiple bond types for each atom pair based on the inner product operation. The multiple bond types can be associated with at least one of single bond, double bond, bond formation, bond collapse, and no change.

[0153] The decoder (420) can generate a probability distribution for each of the plurality of bonding types for each atom pair using the inner product value according to the inner product operation. In addition, the decoder (420) can obtain a transformation matrix that predicts the change in the bonding state of each atom pair using the probability distribution.

[0154] The decoder (420) can generate the probability distribution for the plurality of combination types for each atom pair by applying a softmax function to the inner product values ​​obtained for the plurality of combination types for each atom pair.

[0155] The decoder (420) can generate an atomic characteristic probability distribution of each atom corresponding to the updated embedding (v2) vector using a softmax output layer. Furthermore, the decoder (420) can predict the atomic characteristics of the atoms corresponding to the updated embedding vector using the probability distribution. The atomic characteristics may include charge states of the atoms, which may change during a chemical reaction process of the structure of the plurality of molecules.

[0156] By performing an inner product between the updated embedding vector (v2, or atomic embedding vector, as shown in (a) of Fig. 5b) for the decoder (420) atoms, a score for the potential combined transformation can be calculated, as shown in (b) of Fig. 5. Here, the lowercase letter t can represent a step of the diffusion process.

[0157] This is based on the standard query-key matrix multiplication of the transformer attention mechanism, but in the present invention, the key and the query can share the same projection head. The setting that the key and the query have the same projection head may require multiple projections to accommodate various bond types such as none, single bond, double defect, triple bond, aromatic, etc. This corresponds to the multi-head attention in the transformer, and if expressed in a formula, it can be represented as (c) of Fig. 5b. In this way, the bond prediction can be performed in the Bond prediction heads. This can correspond to the first projection layer (projection layer) 423a illustrated. That is, the bond prediction can be performed in the first projection layer (423a). The first projection layer (423a) can be composed of layers that constitute the bond prediction head.

[0158] Each joint type prediction can be viewed as a head in this multi-head attention scheme, where Softmax regularization is applied to a state such as (d) in Fig. 5b to obtain a distribution of joint transformations for each atom pair. Here, (d) in Fig. 5b may denote a joint transformation.

[0159] In the combined prediction head, a score S is calculated for each combined type k, as shown in (c) of Fig. 5b, where the meaning of each symbol is as shown in (d) of Fig. 5b.

[0160] Meanwhile, in the combined prediction head, the probability for the combined type k of the combined transformation T according to (e) of Fig. 5d can be calculated as shown in (b) of Fig. 5b. In the combined prediction head, the probability distribution of the combined transformation can be obtained by applying softmax regularization as shown in (f) of Fig. 5b.

[0161] In the present invention, each combination type prediction can be understood as a head in a multi-head attention system. The decoder (420) can obtain a normalized probability distribution for the combination transformation of each atom pair by applying softmax normalization to the possible states of the combination transformation according to (e) of FIG. 5b. This can provide an efficient and scalable method for predicting combination transformations directly in the embedding space.

[0162] Next, atom prediction can be performed in the second projection layer (423b) of the decoder (420). The second projection layer (423b) can be composed of layers that constitute an atom prediction head. Since the characteristics of some atoms may change during a chemical reaction, the decoder (420) according to the present invention can perform predictions on atoms in addition to predicting bond transformations between atoms. Atom prediction can be performed in the second projection layer (423b). For this purpose, a standard softmax output layer can be used for each atomic feature. This includes the degree of the atom for each bond type (e.g., the number of single bonds the atom has), and such predictions can help ensure consistency between independent parallel samples of bond transformations and atomic transformations.

[0163] In the atomic prediction head, as illustrated in (c) of Fig. 5c, the probability that the feature Av of the atom v becomes a at step t of the diffusion process can be calculated. Here, (b) of Fig. 5c may represent the embedding of the atom v, and (d) of Fig. 5c may represent the weight matrix of the prediction head for the atomic feature a. In the atomic prediction head, the probability that the atomic feature Av becomes a can be calculated by multiplying the embedding of the atom v (h, see (b) of Fig. 5c)) by the weight matrix (W, see (d) of Fig. 5c) and then applying the softmax function.

[0164] The decoder (420) can obtain a predicted chemical reaction result (425, Predicted Product (Probability Distribution)) predicted from a chemical reaction of a plurality of molecular structures by combining the predicted probability of the combination of atoms and the predicted probability of atoms output from the projection layers (423a, 423b). The predicted chemical reaction result from the projection layers can also be called an initial chemical reaction result.

[0165] In the present invention, the predicted chemical reaction result can be input to a graph diffusion module (430), and the graph diffusion module (430) can sample the initial chemical reaction result using the result of the bond prediction and the result of the atom prediction. The graph diffusion module (430) can stabilize the sampled initial chemical reaction result through a diffusion feedback process. The chemical reaction prediction model can obtain the final chemical reaction result stabilized through the diffusion feedback process.

[0166] In the above diffusion feedback process, the graph diffusion module (430) can repeatedly evaluate each bond transformation of the initial chemical reaction product and remove or change unstable bonds. Furthermore, the graph diffusion module (430) can use the transformation probability matrix and the target transformation matrix to evaluate the predicted bond transformation for each repeatedly performed step in order to predict the bond state change of the initial chemical reaction product, and can generate the final chemical reaction product using an interpolation factor.

[0167] In the present invention, since bond prediction is performed independently across the entire molecular graph, a method is needed to obtain a final, valid molecular structure. For example, during the prediction process, bonds may be predicted to multiple atoms or bonding rules may be violated. To address these issues, a denoising process is required to achieve a valid final state.

[0168] As illustrated in Fig. 5e, a stable final state can be created through multiple sampling and denoising processes (432) for the predicted chemical reaction results.

[0169] FIG. 5e illustrates the diffusion process from the correct data (or target label, ground truth product, 440) to the initial distribution (425), which is based on the initial distribution derived from the dot product matrix of the predicted bond transformations between all atom pairs. Here, the target label corresponding to the correct data may include the actual bonding states and atomic states corresponding to multiple molecular structures input to the encoder. Here, the initial distribution (425) refers to the initial chemical reaction product (425) output from the decoder, which may correspond to a probability distribution. Meanwhile, the diffusion process to the target label is performed during the learning process in the learning unit (440), and in the actual inference stage, a graph diffusion module (430) that has completed learning may be used. In this case, the predicted chemical reaction product (425) may be input to the graph diffusion module (430) that has completed learning, and through a denoising process, the final chemical reaction product may be obtained.

[0170] The learning unit (440) may calculate a loss function between the final chemical reaction result output from the decoder and label data (hereinafter, referred to as “target label”) including actual bonding states and atomic states corresponding to the plurality of molecular structures, and may optimize a parameter of at least one of the encoder and the decoder to minimize the loss function.

[0171] Meanwhile, during the learning process, the target label can be randomly sampled from the interpolated space, interpolated with the initial distribution. Figure 5e shows a noise graph where these random points are denoised into the target label. During prediction, the results sampled from the initial distribution are updated at each step, resulting in a stable final product molecule. A similar process can be applied to atomic features.

[0172] In the diffusion step, the predicted product may correspond to the initial distribution obtained through the dot product matrix of the joint transformation prediction for all atom pairs in the decoder (420), as discussed above. This initial distribution represents the probability distribution of the formation of joints between atoms. The interpolated product is the target label (ground-truth product) interpolated with the initial distribution. In the present invention, a noisy graph can be obtained by sampling random points in this interpolated space. These random points represent the noisy graph that is denoised to the target label. The denoised product denoises the sampled noisy graph so that it gradually gets closer to the target label. This process is performed in multiple steps, and the graph can become more stable at each step. The alternative product means that in some cases, the alternative product may be derived through a different path, which may have a different joint configuration than the target label.

[0173] In this way, in the present invention, the decoder (420) obtains an initial distribution, and the graph diffusion module (430) samples the initial distribution to sample random points. The graph diffusion module (430) interpolates the target label with the initial distribution through interpolation and denoising processes to create an intermediate state. This intermediate state (v3) can be input back to the encoder during the learning process. The graph diffusion module (430) samples random points to obtain a noisy graph, and gradually approaches the target label through a denoising process. This process can be performed repeatedly, and at least one of the encoder, the decoder, and the graph diffusion module can be learned during this repetition.

[0174] The graph diffusion module (430) can be configured to learn how to denoise the predicted chemical reaction product (425), by spreading the target label over the previous output distribution, and then learning how to map the states within the spread interpolation space to the final target label.

[0175] 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 can be obtained from the decoder and derived from the bond prediction head discussed above. This can be expressed as the atom-mapped label of the reaction. The transformation matrix T can identify whether a bond is formed or broken by finding the difference between the input and output adjacency matrices, and the transformation matrix T can be expressed as in (a) of Fig. 5d. 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.

[0176] Meanwhile, the reactants are time-dependent because they continuously change at each diffusion step. This means that the target label also changes at each step. In the present invention, a method for predicting when and where the difference between the reactants and the target label occurred (the reaction center) and, if there is a difference, what changed (the final combined configuration) can be learned, as shown in (b) of Fig. 5d. The parameter at is a randomly sampled interpolation factor that determines the weights between the current transformation matrix and the target transformation matrix. During training, the diffusion process may include obtaining the transformation probabilities Mt from the initial reactant set {Gr} and sampling {Gr(t+1)} from Mt+1 to be used as the input for the next step. Both steps may aim to directly predict the target transformation matrix T. Therefore, the training loss can be defined as shown in (c) of Fig. 5d. During prediction, one can simply encode the input {Gr}, obtain the probability of the combined transformation Mt using the prediction head, and sample from this distribution at each step using standard categorical softmax sampling. In the learning process of the present invention, this process can be repeated continuously and the results can be fed back until a stable configuration is obtained in which the combination no longer changes.

[0177] Meanwhile, in the present invention, the final chemical reaction result can be derived by applying Top-K ranking to a plurality of results stabilized in the diffusion module (430). This can be performed in the Top-K Sampled Product module of the chemical reaction prediction model (400). The Top-K ranking process can be performed in the inference stage. That is, it can be utilized to extract results with a high probability in order to provide the final result to the user. In the present invention, at least one of Top-k sampling and Top-K ranking can be performed, which can be used to obtain more diverse results in the inference process. In Top-k sampling, multiple samples can be generated based on the learned content of the chemical reaction prediction model, and the distribution can be diversified by adjusting the temperature parameter τ. Top-k ranking is used to select the most likely result among the generated samples, calculate the probability of each sample, and select the result with the highest probability as the final prediction.

[0178] In the Top-k sampling process, when sampling the distribution Mt at time step t, the logits can be rescaled by the temperature parameter τ. This can make the softmax distribution depicted in (a) of Figure 6 sharper or more uniform.

[0179] Meanwhile, in order to obtain multiple samples, the present invention can use a temperature schedule τ. In the present invention, when N samples are to be extracted, τi can be set as in (b) of Fig. 6. Here, f(i) is a temperature sampling schedule, and in the present invention, f(i) can be defined as in (c) of Fig. 6. Here, x is a linspace between 0 and 1.

[0180] Next, we examine Top-K ranking. Given sampled candidates, we need a method to rank them by probability. This can be achieved by calculating the probability from the transformation matrix Mt. If we consider the sampled states as sampled tokens for each atom and bond, we can derive the probability for each state from the sampled softmax distribution Mt. This allows us to assign a total probability to the final overall configuration of the molecule (atoms and bond states) and calculate the sampled probability (S). This can be represented as in (d) of Fig. 6. Here, (e), (f), and (g) of Fig. 6 can each be derived from the softmax distribution for atomic features, bond features, and atomic orders. Furthermore, (d) of Fig. 6 can ensure consistency between the predicted atomic order and the actual order calculated from the bonds. In the present 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 probability distribution of the predicted order, and in this way, a high probability score is assigned when the actual order of the generated molecule matches the predicted order.

[0181] Furthermore, since independent sampling from a distribution does not guarantee that all solutions are unique, numerous duplicate samples may result. Therefore, during prediction, the present invention selects the top-k product by taking a simple eigengraph.

[0182] Meanwhile, the above description focuses on Figures 4a and 4b, which include the learning process of a chemical reaction prediction model. However, as previously discussed, the chemical reaction prediction model can have the configurations of Figures 4c and 4d during the inference process. In this case, since everything can be understood identically except for the learning process, the detailed description will be omitted from the above description.

[0183] In the inference process, as described above and illustrated in FIG. 3, a process of receiving information related to a plurality of molecular structures including a first molecular structure and a second molecular structure as inputs of an encoder (S310), a process of obtaining an embedding vector corresponding to the first molecular structure and the second molecular structure using the information related to the plurality of molecular structures in an embedding layer of the encoder (S320), a process of performing an attention operation related to the interaction between atoms of the first molecular structure and the second molecular structure in a multi-head self-attention layer and updating the embedding vector based on the operation (S330), a process of storing the updated embedding vector through the update step in a memory, and inputting the updated embedding vector stored in the memory to a decoder (S340), a process of using the updated embedding vector in the decoder, as a chemical reaction of the first molecular structure and the second molecular structure The final chemical reaction result can be obtained through a process (S350) of performing predicted bond prediction and atom prediction, and a process (S360) of obtaining a final chemical reaction result predicted from a chemical reaction of the first molecular structure and the second molecular structure using the result of the bond prediction and the result of the atom prediction.

[0184] The process for obtaining these final chemical reaction results can be accomplished through the answer generation system (100) illustrated in Fig. 1. At this time, the answer generation system (100) can receive a user query, and the user query can include information about the molecular structure that is the target of chemical reaction prediction.

[0185] In the chemical reaction prediction model (400) according to the present invention, as shown in Fig. 7(a), a chemical reaction result can be obtained as shown in Fig. 7(b) through bond prediction and atomic prediction for a plurality of molecular structures that are the target of a chemical reaction.

[0186] As described above, according to the chemical reaction prediction system and control method thereof according to the present invention, and the learning method of the chemical reaction prediction system, the movement of electrons is modeled through the graph structure of molecules, thereby enabling a good understanding of the chemical reaction mechanism and accurate prediction of the chemical reaction result.

[0187] Furthermore, the chemical reaction prediction system and its control method according to the present invention, as well as the learning method for the chemical reaction prediction system, provide a chemical reaction prediction model that enables a thorough understanding of chemical reaction mechanisms and accurate prediction of chemical reaction results, thereby reducing the time and cost required for experiments by users. This reduces research and development costs and shortens the time to market for new products.

[0188] Meanwhile, according to the chemical reaction prediction system and control method thereof according to the present invention, and the learning method of the chemical reaction prediction system, both the input of the molecular structure and the output of the predicted result are performed in the graph space, thereby enabling forward reaction prediction to be performed regardless of the SMILES permutation and order.

[0189] Furthermore, according to the chemical reaction prediction system and control method thereof according to the present invention, and the learning method of the chemical reaction prediction system, it is possible to simultaneously sample multiple interdependent transformations occurring in parallel within a molecular graph, thereby ensuring consistency between transformations.

[0190] Furthermore, according to the chemical reaction prediction system and control method thereof according to the present invention, and the learning method of the chemical reaction prediction system, in order to solve problems that may arise due to a symmetrical structure, symmetry is broken by including a noise or sampling mechanism, and a valid output structure is formed, thereby preventing the occurrence of an invalid configuration.

[0191] Meanwhile, the present invention discussed above can be implemented 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 can be read by the computer.

[0192] Furthermore, the present invention discussed above can be implemented as computer-readable code or instructions on a program-recorded medium. In other words, the present invention can be provided in the form of a program.

[0193] Meanwhile, computer-readable media include all types of recording devices that store data that can be read by a computer system. Examples of computer-readable media include hard disk drives (HDDs), solid-state disk drives (SSDs), silicon disk drives (SDDs), ROMs, RAMs, CD-ROMs, magnetic tapes, floppy disks, and optical data storage devices.

[0194] Furthermore, the computer-readable medium may include a storage device and may be a server or cloud storage device accessible by an electronic device via communication. In this case, the computer may download the program according to the present invention from the server or cloud storage device via wired or wireless communication.

[0195] Furthermore, in the present invention, the computer described above is an electronic device equipped with a processor, i.e., a CPU (Central Processing Unit), and there is no particular limitation on its type.

[0196] Meanwhile, the detailed description above should not be construed as limiting in any respect and should be considered illustrative. The scope of the present invention should be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the present invention are intended to be included within the scope of the present invention.

Claims

1. A method for predicting chemical reactions performed in cooperation with memory and processor, A step of receiving information related to a plurality of molecular structures as input to an encoder; In the embedding layer of the encoder, a step of obtaining an embedding vector corresponding to the plurality of molecular structures by using information related to the plurality of molecular structures; In a multi-head self-attention layer, a step of performing an attention operation related to the interaction between atoms of the plurality of molecular structures and updating the embedding vector based on the operation; A step of storing the updated embedding vector through the above update step in memory, and inputting the updated embedding vector stored in the memory into a decoder; A step of performing bond prediction and atom prediction predicted as a chemical reaction of the plurality of molecular structures using the updated embedding vector in the decoder; and A chemical reaction prediction method comprising a step of obtaining a final chemical reaction result predicted from a chemical reaction of the plurality of molecular structures by using the results of the above bond prediction and the results of the above atomic prediction.

2. In paragraph 1, The step of obtaining the above chemical reaction result is: A step of sampling the initial chemical reaction results using the results of the above bond prediction and the results of the above atomic prediction; A step of stabilizing the sampled initial chemical reaction result through a diffusion feedback process; and A method for predicting a chemical reaction, characterized in that it comprises a step of obtaining the final chemical reaction result stabilized through the diffusion feedback process.

3. In paragraph 1, Based on the above plurality of molecular structures, the step of obtaining a molecular graph for the plurality of molecular structures by converting atoms into nodes and bonds between atoms into edges is further included. The structure of the above plurality of molecules includes a first molecular structure and a second molecular structure, The step of obtaining the above molecular graph is: A 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 bonding relationships between atoms constituting the first molecular structure into edges using a predetermined graph transformation algorithm; and A chemical reaction prediction method characterized in that it further comprises a step of obtaining the second molecular graph including nodes and edges corresponding to the second molecular structure by converting atoms constituting the second molecular structure into nodes and converting bonding relationships between atoms constituting the second molecular structure into edges using the above-mentioned graph transformation algorithm.

4. In paragraph 3, Information related to the above multiple molecular structures, Includes information about nodes and edges corresponding to the first molecular structure and information about nodes and edges corresponding to the second molecular structure, The above embedding vector is, A method for predicting a chemical reaction, characterized in that it includes at least one piece of information about the atomic type (Type), atomic charge (Charge), number of hydrogens (Hydrogen), number of radical electrons (Radical), and degree (Degree) of a node corresponding to the first molecular structure and a node corresponding to the second molecular structure.

5. In paragraph 4, In the step of updating the above embedding vector, A chemical reaction prediction method characterized in that different biases are added to the attention score calculated by the multi-head self-attention layer according to the type of connection between nodes constituting the first molecular graph and the second molecular graph.

6. In paragraph 5, A method for predicting a chemical reaction, characterized in that the above bond types include single bond types, double bond types, triple bond types, and aromatic bond types.

7. In paragraph 5, Using the nodes and edges corresponding to the first molecular graph and the nodes and edges corresponding to the second molecular graph, Further comprising a step of extracting at least one of an adjacency matrix, a bond type matrix, a shortest paths matrix, and K-hop neighbors corresponding to each of the first molecular graph and the second molecular graph, The above adjacency matrix is, Contains information about direct connections between nodes constituting the first molecular graph and the second molecular graph, The above combination type matrix is, Includes information on the type of coupling between nodes constituting the first molecular graph and the second molecular graph, The above shortest path matrix is, Includes information about the shortest path length between nodes constituting the first molecular graph and the second molecular graph, The above K-level neighbor information is, A chemical reaction prediction method characterized in that each of the nodes constituting the first molecular graph and the second molecular graph includes information on neighboring nodes within K steps.

8. In paragraph 7, In the step of updating the above embedding vector, The output vector of the above multi-head self-attention layer is input to the feed-forward neural network layer, In the feed forward neural network layer, the output vector of the multi-head self-attention layer is updated using at least one of the adjacency matrix, the bond type matrix, the shortest paths matrix, and the K-hop neighbors. A method for predicting a chemical reaction, characterized in that the output vector of the above feedforward neural network layer is specified as the updated embedding vector.

9. In paragraph 1, The above combined prediction is performed by performing a dot-product operation using the updated embedding vector, A chemical reaction prediction method, characterized in that the inner product operation is performed for each vector corresponding to each atom pair of atoms corresponding to the updated embedding vector.

10. In paragraph 9, The step of performing the above combined prediction is: Based on the inner product operation, a step of obtaining an inner product value for each of the plurality of bonding types for each of the atom pairs is included. A method for predicting a chemical reaction, characterized in that the above multiple bond types are related to at least one of single bond, double bond, bond formation, bond collapse and no change.

11. In paragraph 10, The step of performing the above combined prediction is: A step of generating a probability distribution for each of the plurality of bonding types for each of the atom pairs by using the inner product value according to the inner product operation; and A method for predicting a chemical reaction, characterized in that it further comprises a step of obtaining a transformation matrix that predicts a change in the bonding state of each atom pair using the probability distribution.

12. In paragraph 11, In the step of generating the above probability distribution, A chemical reaction prediction method characterized in that, for each of the above atom pairs, a softmax function is applied to the inner product values ​​obtained for the multiple bonding types, thereby generating the probability distribution for the multiple bonding types for each of the above atom pairs.

13. In paragraph 1, In the step of performing the above atomic prediction, A step of generating an atomic characteristic probability distribution of each atom corresponding to the updated embedding vector using a softmax output layer; and Further comprising a step of predicting atomic properties of atoms corresponding to the updated embedding vector using the above probability distribution, A method for predicting a chemical reaction, characterized in that the above atomic properties include the charge states of the atoms that can change during a chemical reaction process of the structure of the plurality of molecules.

14. In paragraph 2, The above diffusion feedback process is, A method for predicting a chemical reaction, characterized in that each bond transformation of the initial chemical reaction product is repeatedly evaluated and unstable bonds are removed or changed.

15. In paragraph 14, The above diffusion feedback process is, A chemical reaction prediction method characterized in that, in order to predict a change in the bonding state of the initial chemical reaction product, a transformation probability matrix and a target transformation matrix are used to evaluate the predicted bonding transformation for each repeatedly performed step, and an interpolation factor is used to generate the final chemical reaction product.

16. A method for predicting chemical reactions performed in cooperation with memory and processor, A step of receiving information related to a plurality of 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 above multiple molecular structures; A step of obtaining an embedding vector corresponding to the molecular graph by using information related to the plurality of molecular structures in the embedding layer of the encoder; In a multi-head self-attention layer, a step of performing an attention operation related to the interaction between atoms of the plurality of molecular structures and updating the embedding vector based on the operation; A step of storing the updated embedding vector through the above update step in memory, and inputting the updated embedding vector stored in the memory into a decoder; A step of 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; A step of obtaining a final chemical reaction result predicted from a chemical reaction of the plurality of molecular structures by using the result of the above bond prediction and the result of the above atomic prediction; A step of calculating a loss function between the final chemical reaction result and the label data including the actual bonding states and atomic states corresponding to the plurality of molecular structures; and A learning method for predicting a chemical reaction, comprising a step of optimizing a parameter of at least one of the encoder and the decoder to minimize the loss function.

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

18. A program stored on a computer-readable medium and executed by one or more processes on an electronic device, The above program is, A step of receiving information related to a plurality of 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 above multiple molecular structures; A step of obtaining an embedding vector corresponding to the molecular graph in the embedding layer of the above encoder; In a multi-head self-attention layer, a step of performing an attention operation related to the interaction between atoms of the plurality of molecular structures and updating the embedding vector based on the operation; A step of storing the updated embedding vector through the above update step in memory, and inputting the updated embedding vector stored in the memory into a decoder; A step of performing bond prediction and atom prediction predicted as a chemical reaction of the plurality of molecular structures using the updated embedding vector in the decoder; and A program including commands for performing a step of obtaining a final chemical reaction result predicted from a chemical reaction of the plurality of molecular structures using the results of the above combination prediction and the results of the above atomic prediction.

Images