Method and system for constructing ultra-large compound library using ai generation model

Abstract

According to an embodiment of the present invention, a compound library construction method performed by at least one processor comprises: a first step of acquiring prior data including first descriptor information indicating molecular structures of compounds; a second step of acquiring training data by converting the first descriptor information of the prior data into second descriptor information; and a third step of training a generation model with the training data and generating new data including new molecular structure information different from the training data, wherein the first descriptor information is in a character string format, and the second descriptor information is in an image format. According to an embodiment of the present invention, the generation model may be trained by applying isomers of the compounds, and as a result, a new molecular structure generated through the generation model may include isomer information.

Description

Method for building a large compound library using an AI generative model and its system

[0001] The present invention relates to a method for constructing a large compound library using an AI-generated model and a system therefor.

[0002]

[0003] Designing and developing new compounds to achieve specific effects or objectives is a crucial process in industries such as the chemical, pharmaceutical, and biotech industries. Traditionally, the process of developing new compounds involves setting specific targets, testing new candidate compounds through experiments, and then building a database of multiple candidates. This approach has been particularly utilized in the field of new drug development.

[0004] However, it is known that developing a new drug takes an average of approximately 15 years and costs 2-3 trillion won. Of this, approximately 6 years are needed to discover a new drug candidate. Once a candidate is discovered, it takes a very long time to determine whether it will cause problems when administered to humans and what pharmacological effects it will have through preclinical and clinical trials. Only through clinical trials can a new drug candidate demonstrate the pharmacological effects that meet the researcher's objectives be determined. Therefore, if the desired pharmacological effect is determined to be absent, a new candidate must be obtained and undergo another clinical trial, resulting in significant costs and time.

[0005]

[0006] Meanwhile, with the recent emergence of various new diseases, the need for rapid drug development to address these emerging diseases is increasing. Therefore, rapid drug development requires the development of new chemical structures that compose the drugs.

[0007]

[0008] In this regard, recent advancements in AI technology, which allows for the generation of novel molecular structures through generative models, are leading to the discovery of new drug candidates. Research is underway to train generative models using the molecular structures of previously known compounds, output new molecular structures, and then derive a group of novel compound candidates based on these models. It is predicted that utilizing AI in new drug development in this manner could shorten the time required for new drug development by approximately 40% and reduce costs by approximately 50%.

[0009]

[0010] Considering the aforementioned research process as an example, a method for deriving compounds used in a specific substance involves training an artificial intelligence (AI) generative model on the structure of a specific substance and compound, then outputting a novel molecular structure that binds the compound to the specific substance. In this case, string data can be used as the training data or output data for the generative model. This string data has the advantage of being easily utilized by various machine learning / deep learning algorithms.

[0011]

[0012] However, string-format data can have problems in reflecting the various properties of molecular structures. In other words, because string-format data is based on molecular formulas, it has limitations in expressing various isomers that have the same molecular formula but different connection patterns or spatial arrangements of constituent atoms within the molecule. Furthermore, even molecules with very similar molecular structures can be expressed as very different strings when expressed as strings, making it difficult to express the similarity between molecules. Furthermore, multiple string data can be generated even when expressing a single molecular formula. That is, even with the same molecular formula, different string data can be derived depending on factors such as which atom is used as the starting point for exploring the molecular structure (DFS) or where the bonds in the ring structure are broken.

[0013]

[0014] Accordingly, the inventor of the present invention recognized the limitations and problems of the generation model to which the above-described string data is applied, and completed the present invention after continuously researching a method and system for constructing a super-large compound library using an AI generation model that can solve the limitations and problems.

[0015]

[0016] While the present invention has been derived at least based on the technical background discussed above, the technical tasks or objectives of the present invention are not limited to resolving the problems or tasks discussed above. That is, in addition to the technical issues discussed above, the present invention can address various technical issues related to the content described below.

[0017]

[0018] (Patent Document 1) Republic of Korea Patent Publication No. 10-2021-0110539

[0019]

[0020] The present invention is intended to derive a library of novel compound candidates that can be used in the chemical, pharmaceutical, and bio industries, and, unlike the generation model described above in the background art, to provide a method and system for expressing and distinguishing isomers of a novel molecular structure candidate compound.

[0021] In addition, the present invention seeks to provide a predetermined preprocessing method for converting dictionary data into training data in order to train a production model of a new compound.

[0022] The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned can be clearly understood by those skilled in the art from the description below.

[0023]

[0024] According to one embodiment of the present invention, a method for constructing a compound library performed by at least one processor includes a first step of obtaining dictionary data including first descriptor information indicating molecular structures of compounds, a second step of converting the first descriptor information of the dictionary data into second descriptor information to obtain learning data, and a third step of training a generation model with the learning data to generate new data including new molecular structure information different from the learning data, wherein the first descriptor information is in a string format and the second descriptor information is in an image format.

[0025] In a compound library construction method according to one embodiment of the present invention, the first descriptor information is characterized in that it includes at least one of a SMILES (Simplified Molecular-Input Line-Entry System) code, a SMARTS (Smiles Arbitrary Target Specification) code, and an InChi (International Chemical Identifier) ​​code.

[0026] In a compound library construction method according to one embodiment of the present invention, the second expression information is characterized in that it is a 2D image or 3D image of a molecular structure.

[0027] In a compound library construction method according to one embodiment of the present invention, the 2D image is characterized in that it represents an isomer selected from the group consisting of structural isomers, stereoisomers, diastereoisomers, and enantiomers.

[0028] The above-described generation model of the compound library construction method according to one embodiment of the present invention is characterized in that it learns a molecular structure based on an image.

[0029] The above-described generation model of the compound library construction method according to one embodiment of the present invention is characterized by performing a step of recognizing the molecular structure of a compound as a set of multiple substructures to generate a tree structure and a step of generating a new molecular structure from the tree structure.

[0030] In a compound library construction method according to one embodiment of the present invention, the first step is characterized by including a step of standardizing a plurality of first expression information according to a preset standard and a step of comparing the standardized plurality of first expression information with each other to remove overlapping first standardization information.

[0031] A compound library construction system according to one embodiment of the present invention comprises a memory and at least one processor connected to the memory and configured to execute at least one computer-readable program included in the memory, wherein the at least one program obtains dictionary data including first descriptor information indicating molecular structures of compounds, converts the first descriptor information of the dictionary data into second descriptor information to generate learning data, trains a generation model with the learning data, and outputs new data including new molecular structure information different from the learning data, wherein the first descriptor information is in a string format and the second descriptor information is in an image format.

[0032]

[0033] The present invention enables the construction of compound libraries more efficiently than conventional techniques. In particular, because it can generate new data by reflecting isomers, the production model can be utilized in various fields, such as new drug development and materials development, to determine the properties of desired final products, such as isomers.

[0034] Furthermore, the present invention increases the possibility of discovering new compounds that were difficult to derive using existing methods. As a result, it can contribute to the discovery of new therapeutics and the development of new functional materials.

[0035] In addition, according to the present invention, a large number of high-quality compound data can be constructed, and thus can be widely used in research activities in various chemical and pharmaceutical fields.

[0036]

[0037] FIG. 1 is a flowchart illustrating a method for constructing a compound library using a generative model according to an embodiment of the present invention.

[0038] FIG. 2 is a diagram showing dictionary data converted into first expressor information according to an embodiment of the present invention.

[0039] FIG. 3 is a diagram showing learning data converted into second expressor information according to an embodiment of the present invention.

[0040] FIG. 4 is a block diagram illustrating a system for constructing a compound library using a generative model according to an embodiment of the present invention.

[0041]

[0042] Various embodiments are now described with reference to the drawings. In this disclosure, various descriptions are provided to facilitate understanding of the present disclosure. However, it will be apparent that these embodiments may be practiced without these specific descriptions.

[0043]

[0044] As used herein, the terms "component," "module," "system," and the like refer to computer-related entities, hardware, firmware, software, a combination of software and hardware, or an execution of software. For example, a component may be, but is not limited to, a procedure running on a processor, a processor, an object, a thread of execution, a program, and / or a computer. For example, both an application running on a computing device and the computing device may be a component. One or more components may reside within a processor and / or a thread of execution. A component may be localized within a single computer. A component may be distributed between two or more computers. Furthermore, these components may execute from various computer-readable media having various data structures stored therein. Components may communicate via local and / or remote processes, for example, by signals comprising one or more data packets (e.g., data from one component interacting with another component in a local system, a distributed system, and / or data transmitted to another system via a network such as the Internet via signals).

[0045] Additionally, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or." That is, unless otherwise specified or clear from context, "X utilizes A or B" is intended to mean either of the natural inclusive permutations. That is, if X utilizes A; X utilizes B; or X utilizes both A and B, "X utilizes A or B" can apply to any of these cases. Furthermore, the term "and / or" as used herein should be understood to refer to and encompass all possible combinations of one or more of the associated items listed.

[0046] Additionally, the terms "comprises" and / or "comprising" should be understood to imply the presence of the features and / or components. However, it should be understood that the terms "comprises" and / or "comprising" do not exclude the presence or addition of one or more other features, components, and / or groups thereof. Furthermore, unless otherwise specified or clear from context to refer to the singular form, the singular in the present disclosure and claims should generally be construed to mean "one or more."

[0047] And, the term "at least one of A or B" should be interpreted to mean "if it includes only A", "if it includes only B", "if it is combined in the composition of A and B".

[0048] Those skilled in the art should further appreciate that the various illustrative logical blocks, configurations, modules, circuits, means, logics, and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate the interchangeability of hardware and software, various illustrative components, blocks, configurations, means, logics, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application. However, such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

[0049] The description of the disclosed embodiments is provided to enable a person skilled in the art to make or use the present disclosure. Various modifications to these embodiments will be apparent to those skilled in the art. The general principles defined herein may be applied to other embodiments without departing from the scope of the present disclosure. Thus, the present disclosure is not limited to the embodiments set forth herein. The present disclosure is to be construed in the widest scope consistent with the principles and novel features disclosed herein.

[0050]

[0051] Hereinafter, specific examples will be described in detail with illustrative drawings. When assigning reference numerals to components in each drawing, it should be noted that, where possible, identical components will be assigned the same reference numerals, even if they appear in different drawings. Furthermore, when describing specific examples, if a detailed description of a related known configuration or function is deemed to hinder understanding of the specific example, such detailed description will be omitted.

[0052]

[0053] The basic process of the present invention involves obtaining dictionary data containing the molecular structures of compounds, converting the dictionary data into training data, and using this data to train an artificial intelligence generative model to acquire new molecular structures. Through this process, a library of new compound candidates is derived.

[0054]

[0055] Specifically, referring to FIG. 1, the present invention relates to a method for constructing a compound library, which is performed by at least one processor, and includes a first step (step S10) of obtaining dictionary data in a first descriptor format representing molecular structures of compounds, a second step (step S20) of converting the first descriptor format of the dictionary data into a second descriptor format to obtain learning data, and a third step (step S40) of generating new data including new molecular structure information different from the learning data by training a generative model with the learning data (step S30). Herein, the first descriptor format may be a string format, and the second descriptor format may be an image format. Each step will be described below.

[0056]

[0057] The first step of the present invention is a step of obtaining dictionary data in the form of a first descriptor representing the molecular structure of compounds.

[0058]

[0059] Dictionary data can be obtained from various sources. For example, public general compound databases, chemical literature, or expert knowledge can be utilized. Additionally, new first descriptor information can be generated using compound synthesis experimental data. Examples of general compound databases include ChEMBL (https: / www.ebi.ac.uk / chembl / ), ZINC database (https: / zinc.docking.org), PubChem (https: / pubchem.ncbi.nlm.nih.gov / ), and Enamine (https: / go.drugbank.com / ). In other words, databases containing data such as molecular structures for various compounds can be used as a source of dictionary data. However, rather than using general compound databases as is regardless of data format, they are used after going through a standardization process to ensure consistency in the first descriptor format. The standardization process for first descriptor information is described in detail below.

[0060]

[0061] Here, the first descriptor format is a string format and may include various expressions representing the molecular structure of the compound. For example, the first descriptor format may include at least one of a Simplified Molecular-Input Line-Entry System (SMILES) code, a Smiles Arbitrary Target Specification (SMARTS) code, and an International Chemical Identifier (InChi) code, but is not limited thereto.

[0062]

[0063] Specifically, SMILES, as an example of a first-order descriptor, can include atoms, bonds, rings, aromaticity, and branches. The process of converting to SMILES based on the above-described elements involves first removing hydrogen atoms from the molecule. Next, the bonds in the molecule's rings are removed one by one, and each ring is numbered. Next, a specific starting atom is determined, and a SMILES string is generated from that starting atom through a depth-first search (DFS). When a branch is encountered, it can be added to the SMILES string.

[0064]

[0065] In the above manner, SMILES expresses different strings depending on the starting point of an atom or the cleavage point of a ring bond even if they have the same molecular formula, so it may be difficult to intuitively reflect isomers that have the same molecular formula but different connection methods or spatial arrangements. Therefore, it is difficult for the first descriptor information such as SMILES to reflect the isomer that determines the physical properties of the compound, and a problem has occurred in which it is difficult to learn the molecular structure or classify new compounds according to the isomer desired by the user in the generative model based on learning and outputting with the first descriptor information.

[0066]

[0067] To address these issues, the present invention includes a second step of acquiring learning data by converting the first descriptor information of the dictionary data acquired in the first step into second descriptor information. Here, the second descriptor information is in image format and visually represents the molecular structure of the compound.

[0068]

[0069] In this way, since the molecular structure can be visually recognized through the second descriptor information, the bond order, bond order, or spatial arrangement between atoms in the compound can be recognized. As a result, isomers can be expressed or identified through the second descriptor information. Specifically, as an example of the second descriptor information, a 2D image can express information on at least one of structural isomers, stereoisomers, diastereoisomers, and enantiomers.

[0070]

[0071] There may be various methods for converting first-expression information into second-expression information. In one embodiment of the present invention, first-expression information (e.g., a SMILES string) may be converted into second-expression information (e.g., a 2D image) through the following process.

[0072]

[0073] First, a SMILES string interpretation step can be performed. In this step, a SMILES string is input. The input SMILES string is converted into a molecule object using a predetermined function or algorithm. Here, a molecule object is data that contains structural information about a chemical molecule, storing information about each atom and bond. Conversion into a molecule object can be interpreted to mean that the SMILES string is interpreted and converted into a data structure that defines the relationships between each atom and bond.

[0074]

[0075] Next, a step can be performed to calculate the 2D coordinates of the molecular structure. This step allows for the appropriate placement of the three-dimensional structure of the chemical molecule on a 2D plane. When the previously converted molecular object is input, the initial 2D coordinates of each atom are calculated using a predetermined function or algorithm. Then, the bond lengths and angles are adjusted to determine the arrangement that minimizes the molecule's energy. After determining the arrangement, the atoms and bonds are positioned so that they do not overlap. Next, a single molecular shape that can be realized in the original three-dimensional structure is maintained.

[0076]

[0077] Finally, an image generation step can be performed. This step converts the chemical structure into an image based on the previously calculated coordinates. The image generation step may include the following substeps: a step of inputting a molecule object with calculated 2D coordinates, a step of rendering a 2D image using a predetermined function or algorithm, and a step of outputting the image using drawing options. Here, the drawing options may include, for example, adjusting the thickness of bond lines, highlighting user-specified atoms or bonds, or adding an atom index image. In addition, the format of the resulting image may be PNG or SVG.

[0078]

[0079] Hereinafter, the standardization process described above will be explained. Since the first descriptor information can be expressed differently depending on the order of starting atoms or the cleavage point even for the same compound, the same compound can be expressed with different first descriptor information. Furthermore, since dictionary data can be obtained from various routes, the expression format may differ depending on each route, so a task to unify them may be necessary. Accordingly, it may be necessary to remove duplicate first descriptor information within the dictionary data. To this end, before performing the second step, a step may be performed to standardize the first descriptor information according to a preset standard, and a step may be performed to compare the standardized first descriptor information with each other to remove duplicate first standardization information.

[0080]

[0081] The standardization process can first read the character data of the chemical structure written in the string label and recognize the format using the format label information. Next, it can be converted according to the standard criteria using a predetermined structure conversion function. Finally, the data converted into the standardized format is stored in a new format.

[0082] In addition, the preset criteria refers to the criteria when the first descriptor information is generated, and may be, for example, at least one element among atomic priority, atomic order, bonding order, stereochemical processing, ring handling, directionality, charge, and isotope. To explain each element more specifically, in the case of atomic priority, atomic number, isotopic mass, stereochemical configuration, hybridization state, etc. are used as criteria, in the case of atomic order, the atom designated as a priority is used as criteria, in the case of bonding order, bonding with the atom designated as a priority is used as criteria, in the case of stereochemical processing, the chiral center is used as criteria, and in the case of ring handling, the number of closed rings may be used as criteria. In addition, in the case of directionality, bonding with aromatic atoms may be used as criteria, and in the case of charge and isotope, explicit charges may be represented as + and -, and isotope may be set as criteria in such a way that the elemental mass is displayed before the atomic symbol.

[0083]

[0084] The third step of the present invention is a step of training a generation model with the training data acquired in the second step to generate new data including new molecular structure information different from the training data.

[0085] Here, the generative model may include an artificial neural network based on artificial intelligence. The artificial neural network includes an input neuron layer, a hidden neuron layer, and an output neuron layer. Each layer includes multiple neurons, and each neuron processes the input data (e.g., represented by x) based on weights (represented by W) and biases (represented by b) and outputs the resulting data (e.g., represented by Wx+b) to the next neuron.

[0086]

[0087] In the above manner, when second-order descriptor information (i.e., an image of a molecular structure) is input to the input neural layer, new data containing a new molecular structure is output from the output neural layer. In the present invention, a deep neural network with a hidden neuron layer depth of 3 or more is used, so that deep learning technology can be applied.

[0088]

[0089] The AI ​​generative model of the present invention can learn the molecular structure of a compound based on an image. The AI ​​generative model is composed of a VAE neural network including an encoder network consisting of an input neuron layer and a hidden neuron layer, and a decoder network consisting of a hidden neuron layer and an output neuron layer. The AI ​​generative model receives molecular structure data from training data and generates and outputs new data with similar molecular structures.

[0090]

[0091] The generative model of the present invention can perform a step of recognizing the molecular structure of a compound as a set of multiple substructures, generating a tree structure, and a step of generating a new molecular structure from the tree structure.

[0092]

[0093] Meanwhile, the generative model can recognize multiple substructures as a set of bonds between atoms, and perform a step of creating a tree structure by combining multiple substructures, and a step of creating a new molecular structure by adding new atoms and bonds between those atoms to the tree structure.

[0094]

[0095] Alternatively, the generative model may recognize the molecular structure of a compound as a set of bonds between atoms, perform a step of creating a tree structure by combining bonds between atoms, and perform a step of creating a new molecular structure by adding new bonds between atoms to the tree structure.

[0096]

[0097] The configuration of the computing device (100) illustrated in FIG. 3 is merely a simplified example. In one embodiment of the present disclosure, the computing device (100) may include other configurations for performing the computing environment of the computing device (100), and only some of the disclosed configurations may constitute the computing device (100).

[0098]

[0099] A computing device (100) may include a processor (110), a memory (120), and a communication interface (130). The processor (110) may be composed of one or more cores, and may include a processor for data analysis and deep learning, such as a central processing unit (CPU), a general purpose graphics processing unit (GPGPU), and a tensor processing unit (TPU).

[0100] The processor (110) can read a computer program stored in the memory (130) and perform data processing for machine learning according to one embodiment of the present disclosure.

[0101] According to one embodiment of the present disclosure, the processor (110) can perform operations for learning a neural network. The processor (110) can perform calculations for learning a neural network, such as processing input data for learning in deep learning (DL), extracting features from the input data, calculating errors, and updating weights of the neural network using backpropagation. At least one of the CPU, GPGPU, and TPU of the processor (110) can process learning of a neural network model. For example, the CPU and GPGPU can together process learning of a neural network model and data classification using a neural network model.

[0102] Additionally, in one embodiment of the present disclosure, processors of multiple computing devices can be used together to process learning of a neural network model and classification of data using a neural network model.

[0103] Additionally, a computer program executed on a computing device according to one embodiment of the present disclosure may be a CPU, GPGPU, or TPU executable program.

[0104]

[0105] The processor (100) can perform the first to third steps illustrated in FIG. 1 using the memory (120). In addition, the generation model described above can also operate through the processor (100) based on the memory (120). Here, the memory (120) can be interpreted in the broadest sense to include volatile memory and non-volatile memory. In addition, the communication interface (130) can connect to an external network to collect preliminary data from a known chemical database or output new data according to an external request. The compound library construction system can be implemented in the form of a computer program through a computing device (100), and since the configuration and execution steps of the system are common to those described above, a detailed description thereof will be omitted.

[0106]

[0107] Although the specific examples described above have been limited to specific examples and drawings, those skilled in the art will recognize that various modifications and variations can be made based on the above teachings. For example, appropriate results can still be achieved even if the described techniques are performed in a different order than described, and / or components of the described systems, structures, devices, circuits, etc. are combined or combined in a different manner than described, or are replaced or substituted with other components or equivalents.

[0108]

[0109] [Explanation of symbols]

[0110] 100: Computer device

[0111] 110: Processor

[0112] 120: Memory

[0113] 130: Communication interface

Claims

1. A method for constructing a compound library, performed by at least one processor, A first step of obtaining dictionary data including first descriptor information representing the molecular structure of compounds; A second step of obtaining learning data by converting the first expressor information of the above dictionary data into second expressor information; and A third step of training a generation model with the above learning data to generate new data including new molecular structure information different from the learning data is included. A method for constructing a compound library, characterized in that the first expressor information is in a string format and the second expressor information is in an image format.

2. In paragraph 1, A method for constructing a compound library, characterized in that the first expression information includes at least one of a SMILES (Simplified Molecular-Input Line-Entry System) code, a SMARTS (Smiles Arbitrary Target Specification) code, and an InChi (International Chemical Identifier) ​​code.

3. In paragraph 1, A method for constructing a compound library, characterized in that the second expression information is a 2D image or 3D image of a molecular structure.

4. In paragraph 3, A method for constructing a compound library, wherein the above 2D image represents an isomer selected from the group consisting of structural isomers, stereoisomers, diastereoisomers and enantiomers.

5. In paragraph 1, A method for constructing a compound library, characterized in that the above generation model learns a molecular structure based on an image.

6. In paragraph 5, The above generative model comprises a step of recognizing the molecular structure of a compound as a set of multiple substructures and generating a tree structure; and A method for constructing a compound library, characterized by performing a step of generating a new molecular structure from the above tree structure.

7. In paragraph 1, The first step is a step of standardizing a plurality of first expressor information according to preset criteria; and A compound library construction method comprising a step of comparing a plurality of standardized first expression information with each other to remove overlapping first standardization information.

8. A computer program stored in a computer-readable recording medium for executing the method according to any one of paragraphs 1 to 7 on a computer.

9. As a compound library construction system, memory; and At least one processor connected to said memory and configured to execute at least one computer-readable program contained in said memory, At least one program above, Obtaining dictionary data including first descriptor information representing the molecular structure of compounds, Convert the first expressor information of the above dictionary data into second expressor information to generate learning data, Train the generative model with the above learning data to output new data including new molecular structure information different from the learning data, A compound library construction system, characterized in that the first expressor information is in string format and the second expressor information is in image format.

Images