Method for generating chemical structure, neural network device and computer readable recording medium

Abstract

The present invention relates to a method for generating a chemical structure, a neural network device, and a computer-readable recording medium. The method for generating a chemical structure implemented by the neural network device includes receiving a target property value and a target structure feature value as input, selecting a first-generation descriptor based on the received input, generating a second-generation descriptor by applying a genetic algorithm to the first-generation descriptor, evaluating the property and structure of each second-generation descriptor by using a first neural network and a second neural network, selecting a candidate descriptor having the target property value and the target structure feature value from the second-generation descriptor based on the evaluation result, and generating a chemical structure for the selected candidate descriptor by using the second neural network.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of Korean Patent Application No. 10-2018-0006275 filed on January 17, 2018 in the Korean Intellectual Property Office, the disclosure of which is incorporated herein in its entirety by reference. Technical Field

[0003] The present disclosure relates to methods and apparatus for generating chemical structures using neural networks. Background Art

[0004] Neural network refers to a computing architecture that imitates the biological brain (modeled on the biological brain). Using advanced neural network technology, various types of electronic systems have been using neural networks to analyze input data and extract effective information.

[0005] In recent years, a lot of research has been conducted on methods for selecting chemical structures to be used in material development by evaluating the properties of chemical structures using neural network technology. In particular, there is a need to develop a method for generating new chemical structures that meet various requirements using neural network technology. Summary of the invention

[0006] A method and apparatus for generating chemical structures using a neural network are provided. In addition, a computer readable medium including a program is provided, which, when executed by a computer, implements the method. The technical problems to be solved are not limited to those as described, but there may be other technical problems.

[0007] Additional aspects will be set forth in part in the description which follows and, in part, will be apparent from the description, or may be learned by practice of the presented embodiments.

[0008] According to an aspect of one embodiment, a method for generating a chemical structure performed by a neural network device includes: receiving a target property value and a target structure feature value as input, selecting a first-generation descriptor based on the received input, generating (multiple) second-generation descriptors by applying a genetic algorithm to the first-generation descriptors, evaluating a property of each second-generation descriptor by using a first neural network and evaluating a structure of each second-generation descriptor by using a second neural network, selecting a candidate descriptor having the target property value and the target structure feature value from the second-generation descriptors based on a result of the evaluation, and generating a chemical structure for the selected candidate descriptor by using the second neural network.

[0009] According to an aspect of another embodiment, a neural network device configured to generate a chemical structure includes a user interface configured to receive a target property value and a target structure feature value as input, a memory storing at least one program, and a processor configured to drive the neural network device by executing the at least one program, wherein the processor selects a first-generation descriptor based on the received input, generates a second-generation descriptor by applying a genetic algorithm to the first-generation descriptors, evaluates a property of each second-generation descriptor by using a first neural network, evaluates a structure of each second-generation descriptor by using a second neural network, and selects a candidate descriptor having the target property value and the target structure feature value from the second-generation descriptors based on a result of the evaluation, and generates a chemical structure for the selected candidate descriptor by using the second neural network.

[0010] According to an aspect of another embodiment, a method for generating a chemical structure by using a conditional variational autoencoder (CVAE) of a neural network device includes converting a first descriptor into a low-dimensional latent variable (latent variable, hidden variable), plotting (mapping) the latent variable into a latent map, selecting from the latent map an area to which the latent variable converted from the first descriptor is not plotted, and converting the latent variable of the selected area into a high-dimensional second descriptor and generating a chemical structure for the second descriptor.

[0011] According to an aspect of another embodiment, a neural network device configured to generate a chemical structure by using a conditional variational autoencoder (CVAE) includes a memory storing at least one program, and a processor configured to drive the neural network device by executing the at least one program, wherein the processor converts a first descriptor into a low-dimensional latent variable, plots the latent variable into a latent map, selects from the latent map a region to which the latent variable converted from the first descriptor is not plotted, and converts the latent variable of the selected region into a high-dimensional second descriptor and generates a chemical structure for the second descriptor.

[0012] According to an aspect of another embodiment, a computer-readable recording medium includes a program that implements the above-described method when executed by a computer. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] These and / or other aspects will become apparent and more readily understood from the following description of the embodiments considered in conjunction with the accompanying drawings, in which:

[0014] Figure 1 A block diagram illustrating a hardware configuration of a neural network device according to an embodiment;

[0015] Figure 2 is a diagram for describing calculations performed by a deep neural network (DNN) according to an embodiment;

[0016] Figure 3 is a diagram for describing calculations performed by a recurrent neural network (RNN) according to an embodiment;

[0017] Figure 4 is a diagram for describing calculations performed by a conditional variational autoencoder (CVAE) according to an embodiment;

[0018] Figure 5 A conceptual diagram for explaining a neural network system for generating chemical structures according to an embodiment;

[0019] Figure 6 is a diagram for exemplarily describing an example of expressing a structural formula of a substance as a character string according to an embodiment;

[0020] Figure 7 is a diagram for describing a process of selecting a first-generation descriptor according to an embodiment;

[0021] Figure 8 is a diagram for describing a process of applying a genetic algorithm to a first generation descriptor according to an embodiment;

[0022] Fig.9A and 9B is a diagram for describing a process of evaluating the properties and structure of a descriptor according to an embodiment;

[0023] Fig.10 is a conceptual diagram of a neural network system configured to generate chemical structures by using CVAE according to an embodiment;

[0024] Fig.11 is a flowchart for describing a method for generating a chemical structure by using a DNN and a RNN according to an embodiment; and

[0025] Fig.12 is a flow chart for describing a method of generating a chemical structure by using CVAE according to an embodiment. DETAILED DESCRIPTION

[0026] Embodiments will now be described in detail, examples of which are shown in the accompanying drawings, wherein the same reference numerals refer to the same elements throughout. In this regard, the present embodiment may have different forms and should not be construed as being limited to the description set forth herein. Therefore, the following only describes the embodiments to illustrate aspects by reference to the accompanying drawings. As used herein, the term "and / or" includes any and all combinations of one or more of the related enumerated items. Statements such as "at least one (kind) of..." modify the entire list of elements when before or after the list of elements without modifying the individual elements of the list.

[0027] The terms "according to some embodiments" or "according to (an) embodiment" used throughout the specification do not necessarily refer to the same embodiment.

[0028] Some embodiments of the present disclosure can be described by function block (frame) construction and various processing operations. Some or all of these function blocks can be implemented using various numbers of hardware and / or software components (components, ingredients) that perform specific functions. For example, the function blocks of the present disclosure can be implemented using one or more microprocessors or circuits for a given function. In addition, for example, the function blocks of the present disclosure can be implemented in a variety of programming or scripting languages. The function blocks can be implemented using algorithms running on one or more processors. The present disclosure can also be used in conventional techniques for electronic construction, signal processing, and / or data processing. The terms "mechanism", "element", "unit" and "construction" can be used in a broad sense and are not limited to mechanical and physical constructions.

[0029] In addition, the connecting lines or connecting members between components shown in the drawings only illustrate functional connections and / or physical or circuit connections. In actual equipment, the connections between components can be provided by various functional connections, physical connections, or circuit connections that can be replaced or added.

[0030] Meanwhile, regarding the terms used herein, a descriptor as data used in a neural network system refers to an indicative value used to describe a property (feature) of a substance and can be obtained by performing a relatively simple calculation on a given substance. According to an embodiment, the descriptor may include a molecular structure fingerprint (e.g., a Morgan fingerprint and an extended connectivity fingerprint (ECFP)) indicating whether a specific local structure is included and a quantitative structure-property relationship (QSPR) configured with a value that can be immediately calculated, such as a molecular weight or the number of local structures (e.g., a ring) included in a molecular structure.

[0031] In addition, properties refer to characteristics possessed by a substance and may be real numbers measured by experiments or may be calculated by simulation. For example, when the substance is a display (display) material, the property may be a transmission wavelength, an emission wavelength, etc. for light. When the substance is a battery material, the property may be voltage. Unlike descriptors, calculation of properties may require complex simulations and take a long time.

[0032] In addition, structure refers to the atomic level structure of a substance. In order to derive (derive) properties by performing first principle calculations, the structure needs to be expressed at the atomic level. Therefore, the structure of a substance needs to be derived to the atomic level to generate a new chemical structure. The structure can be a structural formula based on atomic bonding relationships or a string in a simple format (one-dimensional). The format of the string expressing the structure can be a simplified molecular linear input system (specification) (SMILES) code, a SmilesArbitrary Target Specification (SMARTS) code, an International Chemical Identification (InChi) code, etc.

[0033] In addition, factors refer to elements that define the relationship between descriptors, properties, and structures. Factors can be determined by machine learning based on descriptor-property-structure formulas stored in a database. Therefore, it is possible to find out how factors are associated with descriptors, properties, and structures.

[0034] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings.

[0035] Figure 1 A block diagram illustrating a hardware configuration of a neural network device according to an embodiment.

[0036] The neural network device 100 may be implemented using various types of devices such as personal computers (PCs), server devices, mobile devices, and embedded devices. Examples of the neural network device 100 may include, but are not limited to, smartphones, tablet devices, augmented reality (AR) devices, Internet of Things (IoT) devices, autonomous vehicles, robots, medical devices, and the like that use neural networks to perform speech recognition, image recognition, image classification, and the like. In addition, the neural network device 100 may be a dedicated hardware (HW) accelerator installed on the devices described above. The neural network device 100 may be a hardware accelerator such as a neural processor (processing unit) (neural network processor) (NPU), a tensor processing unit (TPU), and a neural engine as a dedicated module (component) for driving a neural network, but is not limited thereto.

[0037] Reference Figure 1 , the neural network device 100 includes a processor 110 and a memory 120 . Figure 1 Only the components of the neural network device 100 related to the embodiments of the present disclosure are described. Therefore, it is clear to those skilled in the art that the neural network device 100 may further include Figure 1 Any other generic components other than those shown in .

[0038] The processor 110 controls the overall (overall) function for driving the neural network device 100. For example, the processor 110 controls the overall operation of the neural network device 100 by executing a program stored in the memory 120 of the neural network device 100. The processor 110 may be implemented as a central processing unit (CPU), a graphics processing unit (GPU), an application processor (AP), etc. provided in the neural network device 100, without being limited thereto.

[0039] Memory 120 is a hardware component that stores various data processed in neural network device 100. For example, memory 120 may store data processed by neural network device 100 and to be processed by neural network device 100. Memory 120 may also store applications, drivers, etc. to be driven by neural network device 100. Memory 120 may include random access memory (RAM) such as dynamic random access memory (DRAM) and static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), CD-ROM, Blu-ray, or other optical disk storage, hard disk drive (HDD), solid state drive (SSD), or flash memory (flash memory).

[0040] At the same time, the memory 120 stores the structural feature values ​​and the property values ​​that match each other as a group. The neural network device 100 can read the structural feature values ​​and the property values ​​from the memory 120 or write the structural feature values ​​and the property values ​​into the memory 120. The property value refers to the characteristics of a substance and can be a real number measured by an experiment or can be calculated by simulation. In addition, the structural feature value as an indicator value for expressing the structure of a substance can be a molecular structure fingerprint (e.g., Morgan fingerprint and ECFP) indicating whether a specific local structure is included.

[0041] The processor 110 may drive a deep neural network (DNN), a recurrent neural network (RNN), and a conditional variational autoencoder (CVAE).

[0042] The processor 110 may allow the DNN to learn by using the descriptors and properties (property values) and may determine factors defining the relationship between the descriptors and the properties in the process. By driving the trained DNN, the processor 110 may perform calculations using the descriptors as input data to the input layer and generate property values ​​as output data based on the calculation results.

[0043] The processor 110 may allow the RNN to learn by using the descriptor and the structure (structural feature value) and may determine the factors defining the relationship between the descriptor and the structure in the process. By driving the trained RNN, the processor 110 may perform calculations using the descriptor or factor as input data to the input layer and generate structural feature values ​​as output data based on the calculation results.

[0044] The processor 110 can generate a new chemical structure that does not exist in the database by using CVAE. In particular, a descriptor of high-dimensional data stored in the database can be used as input data to the input layer of CVAE, and the high-dimensional descriptor can be converted into a latent variable as low-dimensional data by an encoder. Then, the low-dimensional latent variable is decoded by a decoder of CVAE. Finally, a descriptor of high-dimensional data corresponding to the new chemical structure can be output from the output layer.

[0045] Meanwhile, the neural network device 100 may further include a user interface (not shown). The user interface refers to a device for inputting data to control the neural network device 100. Examples of the user interface may include, but are not limited to, a keypad, a dome switch, a touch pad (e.g., a capacitive superposition type, a resistive superposition type, an infrared beam type, a surface acoustic wave type, an integral strain gauge type, and a piezoelectric type), a jog wheel, and a jog switch.

[0046] Hereinafter, a method of generating a chemical structure and evaluating the generated chemical structure by using the neural network device 100 will be described in detail. The method to be described below may be performed by the processor 110 and the memory 120 of the neural network device 100.

[0047] Figure 2 is a diagram for describing calculations performed by a DNN according to an embodiment.

[0048] Reference Figure 2 The DNN 20 has a structure including an input layer, a hidden layer, and an output layer. 1 and I 2 ) performs calculations and generates output data (e.g., 1 and O 2 ).

[0049] For example, Figure 2As shown in FIG. 1 , DNN 20 may include an input layer (layer 1), two hidden layers (layers 2 and 3), and an output layer (layer 4). Since DNN 20 may include many layers for processing effective information, DNN 20 may process complex data compared to a neural network including a single layer. Meanwhile, although Figure 2 The DNN 20 shown in FIG. 1 includes 4 layers, but the DNN 20 is only an example and may also include more or fewer layers and more or fewer channels than those shown therein. That is, the DNN 20 may have Figure 2 The various layer structures shown in FIG.

[0050] Each of the layers included in the DNN 20 may have a plurality of channels. The channels may correspond to a plurality of artificial nodes, which are referred to as neurons, processing elements (PEs), units, or similar terms. For example, Figure 2 As shown in FIG, layer 1 may include two channels (nodes), and layers 2 and 3 may include three channels, respectively. However, the layers are merely examples, and the layers included in DNN 20 may each have various numbers of channels (nodes).

[0051] Channels included in each of the layers of the DNN 20 may be interconnected to process data. For example, a channel may perform calculations on data received from a channel of one layer and output calculation results to a channel of another layer.

[0052] The input and output of each channel may be referred to as input activation and output activation. That is, the activation may not only be the output of one channel, but also a parameter corresponding to the input of the channel included in the successive layer. At the same time, each of the channels may determine its activation based on the activation and weight received from the channel included in the previous layer. The weight is a parameter for calculating the output activation of each channel and may be a value assigned to the relationship between the channels.

[0053] Each of the channels can be processed by a computational unit or processing element that receives an input and outputs an activation output. The input-output of each channel can be plotted. For example, when σ is an activation function, is the weight from the kth channel included in the (i-1)th layer to the jth channel included in the i-th layer, is the bias of the jth channel included in the i-th layer, and When the activation is the jth channel of the i-th layer, the activation can be calculated using the following expression 1

[0054] Expression 1

[0055]

[0056] like Figure 2As shown in FIG. 1 , the activation of the first channel CH1 of the second layer 2 can be performed as In addition, according to expression 1, Possible However, the above expression 1 is only an example for describing the activation and weight for processing data in the DNN 20, and the embodiment is not limited thereto. The activation may be a value obtained by inputting the sum of activations received from the previous layer to the activation function and processing the result with a rectified linear unit (ReLU).

[0057] According to an embodiment, DNN 20 may determine factors defining the relationship between descriptors and properties by learning using descriptors and property values. That is, among layers 1 to 4 constituting DNN 20, descriptors correspond to the input layer (layer 1), property values ​​correspond to the output layer (layer 4), and factors correspond to at least one hidden layer (layers 2 and 3).

[0058] The DNN 20 may perform calculations using the descriptors as input data to an input layer, and generate property values ​​as output data based on the calculation results.

[0059] Figure 3 is a diagram for describing calculations performed by an RNN according to an embodiment.

[0060] In the following, for the convenience of description, the above reference will not be repeated. Figure 2 Description given.

[0061] The RNN 30 is a neural network that learns data that changes over time, such as time series data, and is constructed by connecting a reference time point t and a next time point t + 1. That is, the RNN 30 is a neural network in which the time aspect is taken into account and is capable of efficiently learning patterns from data that is input sequentially or data that is input in a feature sequence by modifying the model to allow recursive input to the hidden layer of the neural network.

[0062] Reference Figure 3 , showing nodes s constituting the hidden layer of RNN 30. Node s can perform calculations based on input data x and generate output data o. RNN 30 can apply the same task to all sequences, and the final output result of node s is affected by the results of previous calculations.

[0063] RNN 31 is an unfolded RNN 30 with loops. The term "unfolded" RNN 30 refers to expressing RNN 30 for the entire sequence. In RNN 31, x t is the input value at time step t, and s t is the hidden state at time step t. tIt can be expressed by the following expression 2. In expression 2, a hyperbolic tangent or a Relu function can be used as the function f. The s used to calculate the first hidden state t-1 can usually be initialized to 0. In addition, in RNN 31, o t is the output value at time step t.

[0064] Expression 2

[0065]

[0066] Here, s t is the memory part of the network and stores information about events at previous time steps. Output value o t Depends only on the storage of the current time step t.

[0067] Meanwhile, unlike the existing neural network structure in which parameters are different from each other, RNN 31 shares parameters U, V, and W for all time steps. That is, since each step of RNN 31 performs almost the same calculation except for the input value, the number of parameters to be learned can be reduced.

[0068] According to an embodiment, the RNN 32 may determine factors defining the relationship between the descriptor and the structure by learning using the descriptor and the structure (structure feature value). Figure 2 As described, the factor may be at least one hidden layer. The RNN 32 may perform calculations using the descriptors or factors as input data to the input layer and generate structural feature values ​​as output data based on the calculation results.

[0069] For example, when the string representing the structural feature value is "ABCDEFG", the input and output of each time step may be "ABC", "BCD", "CDE", "DEF", and "EFG". That is, each step of the RNN can use the next string of the string input at time point t as the input at time point t+1.

[0070] The RNN 32 may perform learning to maximize the probability of generating a string of structure s representing data h and s (ie, factor and structure data). The probability p of generating a string may be expressed by the following Expression 3.

[0071] Expression 3

[0072]

[0073] When the construction of the RNN part is completed by learning (i.e., when the parameters of the RNN model for determining the structural factor are determined), the factor obtained by encoding the descriptor can be decoded. According to another embodiment, the decoding is performed sequentially as follows: the first part ABC of the decoded string is obtained by using the factor e(x) as input, the part "BCD" of the string is obtained using the obtained part "ABC" as input, and the part "CDE" of the string is obtained using the obtained part "BCD".

[0074] Reference Figure 3 , the first two characters of the portion of the string at time point t are the same as the last two characters of the portion of the string at time point t-1. Then, the characters at each time point are grouped to output a single string ABCDEFG.

[0075] Figure 4 is a diagram for describing calculations performed by CVAE according to an embodiment.

[0076] Reference Figure 4 , CVAE 40 has a structure including an input layer, an encoder, a decoder, and an output layer. High-dimensional data such as an image stored in a database is used as input data to the input layer of CVAE 40, and the encoder performs encoding to convert the high-dimensional data into a low-dimensional latent variable z. According to an embodiment, the latent variable z follows a normal distribution with a mean value μ and a variance σ (Σ) and can be two to fifty dimensional data. Then, the decoder decodes the low-dimensional latent variable z to output a new image (high-dimensional data) that does not exist in the database from the output layer.

[0077] For example, when using images of human shapes as input data, the latent variables may be the shape of the subject, the viewpoint of the camera, information about the light source, etc. When using images of numbers as input data, the latent variables may be the angle between lines, the aspect ratio, etc.

[0078] Meanwhile, the latent variable z may be mapped into a latent map. When a predetermined value z' included in a region of the latent map to which the latent variable z is not mapped is input to the decoder, new data not stored in the database may be generated from the output layer.

[0079] In the CVAE 40, a descriptor as high-dimensional data stored in a database may be used as input data to an input layer, and encoding of converting the high-dimensional descriptor into a low-dimensional latent variable may be performed by an encoder. Then, the low-dimensional latent variable may be decoded by a decoder of the CVAE 40. Finally, a descriptor as high-dimensional data corresponding to a new chemical structure may be output from an output layer.

[0080] Figure 5A conceptual diagram illustrating a neural network system for generating chemical structures according to an embodiment.

[0081] Reference Figure 5 , illustrating a neural network system configured to generate chemical structures by using a DNN 500, a RNN 510, and a CVAE 520.

[0082] Descriptors of data used in the neural network system may include QSPR descriptors configured with immediately computable values ​​such as a molecular structural fingerprint (ECFP) indicating whether a specific local structure is included. Properties refer to characteristics possessed by a substance and may be real numbers measured experimentally or may be calculated by simulation. Structure refers to the structure of a substance at the atomic level. The string format for expressing the structure may be simplified as a SMILES code, a SMARTS code, an InChi code, etc. For example, a structural formula may be expressed by a SMILES code as shown in Expression 4 or by a SMARTS code as shown in Expression 5.

[0083] Expression 4

[0084] OC1=C(C=C2C=CNC2=C1)C1=C(C=CC=C1)C1=CC2=C(NC=C2)C=C1

[0085] Expression 5

[0086] [#6]-1-[#6]=[#6]-[#6]-2=[#6]-1-[#6]-1=[#6](-[#6]-[#6]=[#7]-[#6]-1)-[#7]=[#6]-2

[0087] A factor is an element defining the relationship between a descriptor, a property, and a structure. A factor may be at least one hidden layer. When a factor includes a plurality of hidden layers, a factor defining the relationship between a descriptor and a property, a factor defining the relationship between a descriptor and a structure, etc. may be determined for each hidden layer.

[0088] A potential map is a visualized (visualizable) coordinate to which two-dimensionally or three-dimensionally expressed descriptors are plotted. Similarities between descriptors can be visualized and identified by using a potential map.

[0089] DNN 500 can perform calculations using descriptors as input data to the input layer and generate properties (property values) as output data based on the calculation results. RNN 510 can perform calculations using descriptors or factors as input data to the input layer and generate structures (structural feature values) as output data based on the calculation results. As a string format for expressing structural feature values, SMILES code, SMARTS code, etc. can be used.

[0090] In addition, the DNN 500 and the RNN 510 can determine the factors defining the relationship between the descriptor, the property, and the structure by performing learning based on the relationship between the property (property value) and the structure (structure feature value) stored in the memory. According to an embodiment, the factor may be at least one hidden layer. When the factor includes a plurality of hidden layers, the factors defining the relationship between the descriptor and the property and the factors defining the relationship between the descriptor and the structure can be determined for each hidden layer.

[0091] The RNN 510 may perform calculations using the factors determined by the DNN 500 as input data to an input layer and generate structural feature values ​​as output data based on the calculation results.

[0092] In CVAE 520, a descriptor as high-dimensional data stored in a database is used as input data to an input layer, and encoding of the high-dimensional descriptor into a latent variable as low-dimensional data can be performed by an encoder. Then, when the decoder of CVAE 520 decodes the low-dimensional latent variable, a descriptor as high-dimensional data corresponding to a new chemical structure can be output from the output layer. In the case where the descriptor output from CVAE 520 is used as input data of DNN 500, a property value can be generated. In the case where the output descriptor is used as input data of RNN 510, a structural feature value can be generated.

[0093] Figure 6 FIG. 2 is a diagram for exemplarily describing an example of expressing a structural formula of a substance as a character string according to an embodiment.

[0094] Reference Figure 6 , illustrating an example of expressing the structural formula of a substance as a character string. In the entire structure 610 of a substance, there may be a local structure 620 that is to be maintained, ie, must be included in the chemical structure to be finally generated.

[0095] The local structure 620 can be expressed by a character string such as a SMILES code, a SMARTS code, or an InChi code. For example, the local structure 620 can be expressed as a SMARTS code as shown in Expression 6.

[0096] Expression 6

[0097] [#6]-1-[#7]=[#6]-[#6]-2-[#6]-1=[#6]-[#6]-1=[#6]-[#6]=[#6]-[#7]=[#6]-2-1

[0098] In addition, when changes in bonds and atomic parts are allowed in the local structure 620, the local structure 620 can be expressed as a SMARTS code 621. When changes in atomic parts are allowed in the local structure 620, the local structure 620 can be expressed as a SMARTS code 622. When bonds and atomic parts are fixed in the local structure 620, the local structure 620 can be expressed as a SMARTS code 623.

[0099] Figure 7 is a diagram for describing a process of selecting a first-generation descriptor according to an embodiment.

[0100] Reference Figure 7 , the neural network device 700 may receive a target property value 710 and a target structure feature value 720 as input values. The neural network device 700 may generate a chemical structure satisfying the target property value 710 and the target structure feature value 720. That is, the neural network device may generate a chemical structure including a predetermined local structure and having improved properties.

[0101] The target property value 710 refers to the value of a certain property possessed by the chemical structure finally generated by the neural network device 700. According to one embodiment, the target property value 710 may be a refractive index value, an elastic modulus, a melting point, a transmission wavelength, and an emission wavelength. For example, the neural network device 700 may receive "emission wavelength: λ=350 nm" as the target property value 710.

[0102] According to another embodiment, the target property value 710 may be in an increasing (+) direction or a decreasing (-) direction instead of a given numerical value. The memory 730 stores the structural feature values ​​and the property values ​​that match each other as a group. The neural network device 700 may select a given property value stored in the memory 730. In this case, since the neural network device 700 sets the target property value 710 as an increasing direction instead of a certain numerical value, the chemical structure that will be finally generated may have a property value greater than the property value selected in the memory 730.

[0103] The target structure feature value 720 refers to the value of the local structure that must be included in the chemical structure that will be finally generated by the neural network device 700. The local structure can be expressed as a SMILES code or a SMARTS code. In addition, in order to numerically express the local structure expressed as a SMILES code or a SMARTS code, descriptors such as ECFP and QSPR can be used.

[0104] According to an embodiment, the target structure feature value 720 may be a descriptor of a specific chemical structure. For example, the local structure 721 may be expressed as a SMILES code 722a or a SMARTS code 722b. The neural network device 700 may receive a descriptor 723 corresponding to the SMILES code 722a or the SMARTS code 722b as the target structure feature value 720. For ease of understanding, Figure 7 The SMILES code and SMARTS code corresponding to the descriptor 723 and the chemical structure are described.

[0105] Meanwhile, the descriptor may be a binary value having a length of 100 bits. Hereinafter, for the convenience of description, the descriptor will be expressed only by the bit (digit) having the bit value (bit value) of "1". For example, the descriptor 723 may be expressed by [0 2 4 8 9 10 18 19 20 29 30 42 47 48 49 51 53 56 57 63 64 66 7279 83 86 87 93 95 97 98]. In this case, the bits 0, 2, 4, 8, 9, 10, etc. of the descriptor 723 indicate the bit value "1" and the bits 1, 3, 5, 6, 7, 11, etc. indicate the bit value "0".

[0106] The neural network device 700 may select a first-generation descriptor 740 based on the received target property value 710 and the target structural feature value 720. In particular, the neural network device 700 may select a first-generation descriptor 740 having a property value similar to the received target property value 710 and a structural feature value similar to the received target structural feature value 720 from the memory 730 in which structural feature values ​​and property values ​​matching each other are stored.

[0107] Reference Figure 7 , the neural network device 700 receives an input of "emission wavelength: λ=350nm" as a target property value 710 and an input of a descriptor 723 as a target structural feature value 720. The neural network device 700 may select a chemical structure having a structural feature value and a property value similar to the received input from the memory 730. The selected chemical structure may have a target structural feature value 740 expressed as descriptors 1 and 2, and a plurality of descriptors (e.g., descriptors 1 and 2) included in the target structural feature value 740 correspond to the first generation descriptors. At the same time, for ease of understanding, Figure 7 The SMILES code, SMARTS code, and chemical structure corresponding to descriptors 1 and 2, respectively, are described.

[0108] Figure 8 is a diagram for describing a process of applying a genetic algorithm to a first generation descriptor according to an embodiment.

[0109] Reference Figure 8 , the neural network device can generate a second generation of descriptors by applying a genetic algorithm to the first generation of descriptors. A genetic algorithm is an algorithm for generating new candidate substances by performing crossover (overlap) and mutation operations (runs) on previous candidate substances. The operation of a genetic algorithm is an arbitrary process. A genetic algorithm includes operations such as selection, crossover, and mutation. Selection is a process for selecting candidate substances to be passed from one generation to the next. Crossover is a process for generating candidate substances of the next generation via pairing within a generation. Mutation is a process for changing a candidate substance via random changes in the order or value of descriptors in a candidate substance.

[0110] Reference Figure 8 , the neural network device can select the first generation descriptor. The first generation descriptor is descriptors 1 to 4, and descriptors 1 to 4 can be expressed by binary codes.

[0111] The neural network device may perform crossover and mutation operations on the descriptors 1 to 4 selected as the first generation descriptors. As a result of the crossover and mutation operations, the neural network device may generate second generation descriptors from the first generation descriptors (descriptors 1 to 4).

[0112] Fig.9A and 9B is a diagram for describing a process of evaluating the properties and structure of a descriptor according to an embodiment.

[0113] Reference Fig.9A , the DNN 900 of the neural network device is a neural network in which factors defining the relationship between a descriptor and a property are determined through learning using the descriptor and the property (property value).

[0114] In particular, DNN 900 learns an encoding function e(·) and a prediction function f(·) based on (x, y) data (there may be more) indicating descriptors and property values. The encoding function e(·) is a function used to derive hidden parameters from descriptors. When the descriptor is x, the factor h (= e(x)) about the descriptor x can be determined via the encoding function. The prediction function f(·) is a function used to predict properties based on factors obtained on the basis of encoding. In the case where the property is y, when the factor h is substituted into the prediction function, y=f(h). In the learning operation, DNN 900 may determine the factor to maximize the probability of the data (x, y) or minimize the difference between the predicted property value f(e(x)) and the true property y.

[0115] DNN 900 performs calculations using descriptors as input data to an input layer and generates property values ​​as output data based on the calculation results. The factors may be undisclosed core elements for representing the relationship between the descriptors and the properties. Among the multiple layers constituting DNN 900, the descriptors may be input layers, the factors may be hidden layers, and the property values ​​may be output layers.

[0116] The neural network device inputs the second-generation descriptor to the DNN 900 as input data and generates a property value of each (corresponding) second-generation descriptor as output data by applying the encoding function e(·) and the prediction function f(·) thereto. The neural network device can evaluate the property of the second-generation descriptor by comparing the generated property value with the target property value.

[0117] In addition, the RNN 910 of the neural network device is a neural network in which factors of the relationship between the descriptor and the structure are defined through learning using the descriptor and the structure (structure feature value).

[0118] In particular, the RNN 910 learns the encoding function e(·) and the decoding function d(·) based on the (x, s) data indicating the descriptor and the structural feature value, and can determine the factor in the process of learning the decoding function d(·). At the same time, when the encoding function e(·) is determined through the learning process of the DNN 900, the RNN 910 can learn the decoding function d(·) based on the (h, s) data as the factor and the structural feature value by using the encoding function e(·) determined by the DNN 900.

[0119] The decoding function d(·) is a function for generating the structure of a substance at an atomic level by using factors obtained from the descriptor. When the structure is expressed as s, s=d(h) is established.

[0120] RNN 910 performs calculations using the descriptors as input data to the input layer and generates structural feature values ​​as output data based on the calculation results. The factors may be undisclosed core elements for representing the relationship between the descriptors and the structures and may be hidden layers of RNN 910. Hidden layers generated by the learning process of DNN 900 may also be used.

[0121] The neural network device inputs the second-generation descriptor to the RNN 910 as input data and generates a structural feature value of each second-generation descriptor as output data by applying an encoding function e(·) and a decoding function d(·) thereto. The neural network device can evaluate the structure of the second-generation descriptor by identifying whether the generated structural feature value includes a target structural feature value.

[0122] In particular, the RNN 910 may generate a structural feature value (final chemical structure) in the form of a SMILES code. When the target structural feature value (local structure) is input in the form of a SMARTS code, the neural network device may determine whether the structural feature value includes the target structural feature value by comparing the SMILES code corresponding to the structural feature value with the SMARTS code corresponding to the target structural feature value using the RDkit library. When the structural feature value generated by the RNN 910 includes the target structural feature value, that is, when the generated final chemical structure includes a local structure, the neural network device may select the second generation descriptor input to the RNN 910 as a candidate descriptor. Meanwhile, the method of expressing the structural feature value (final chemical structure) and the target structural feature value (local structure) is not limited thereto.

[0123] The neural network device may evaluate the properties of the second generation descriptors by using the DNN 900 and the structures of the second generation descriptors by using the RNN 910. In addition, the neural network device may select a candidate descriptor having a target property value and a target structure feature value from the second generation descriptors based on the evaluation results of the properties and structures. Finally, the neural network device generates a chemical structure for the selected candidate descriptors, thereby generating an improved chemical structure including the specific local structure and having the target property value.

[0124] Meanwhile, when the property value and / or the structural feature value of the second generation descriptor cannot reach the target value, the neural network device may generate a third generation descriptor by applying a genetic algorithm to the second generation descriptor. The neural network device may evaluate the property of the third generation descriptor by using the DNN 900 and the structure of the third generation descriptor by using the RNN 910. The neural network device may develop the descriptor by using the genetic algorithm until a descriptor that satisfies the target property value and the target structural feature value is generated.

[0125] In addition, when the number of candidate descriptors having the target property value and the target structure feature value is less than a preset value as a result of the evaluation of the property and structure of the second generation descriptor, the neural network device may change a transformation index applied to the genetic algorithm.

[0126] Meanwhile, the transformation degree of the first generation descriptor may be changed according to the transformation index applied to the genetic algorithm. According to an embodiment, reducing the transformation index may refer to reducing the number of operations with a large transformation degree of the first generation descriptor among the selection, crossover, mutation, and replacement operations constituting the genetic algorithm. However, the method of changing the transformation index is not limited thereto.

[0127] The neural network device may regenerate (regenerate) the second generation descriptor from the first generation descriptor after changing the transformation index of the genetic algorithm. According to an embodiment, when 100 second generation descriptors are generated and the number of descriptors having the target property value and the target structural feature value is 10 or less, the neural network device may reduce the transformation index of the genetic algorithm and then apply the genetic algorithm to the first generation descriptor again to regenerate the second generation descriptor.

[0128] Reference Fig. 9B The neural network device can generate a chemical structure including a specific local structure and having improved properties by developing a descriptor by using a genetic algorithm, and evaluating a property value and a structural feature value of the generated descriptor by using a DNN 900 and a RNN 910.

[0129] According to an embodiment, the neural network device may gradually (step by step) generate a chemical structure including a specific local structure and having improved properties. For example, the neural network device may generate a chemical structure including the local structure 921 and an increased wavelength (T1 value).

[0130] In particular, the neural network device may select a chemical structure 922 including a local structure 921, having a T1 value of 355.93, and a structural feature value of "OC(=O)C1=CC=CC=C1NC1=CC=CC=C1". The neural network device may generate a chemical structure 923 including the local structure 921 and having an increased T1 value of 420.11 by applying a genetic algorithm to the structural feature value "OC(=O)C1=CC=CC=C1NC1=CC=CC=C1" of the selected chemical structure 922. The generated chemical structure 923 has a structural feature value of "N(C1=CC=CC=C1)C1=C2C=CC=CC2=NC2=C1C=CC=C2".

[0131] In this way, the neural network device can generate a chemical structure including the local structure 921 and having an increased T1 value by applying the genetic algorithm to the generated chemical structure 923. The neural network device can generate a chemical structure by applying the genetic algorithm thereto until a chemical structure having a property value similar to a target property value is generated.

[0132] According to another embodiment, the neural network device may also generate a chemical structure including a specific local structure and a reduced T1 value by applying a genetic algorithm to the generated chemical structure 923 .

[0133] Fig.10 is a conceptual diagram of a neural network system configured to generate chemical structures by using CVAE according to an embodiment.

[0134] The CVAE 1000 of the neural network device uses a descriptor, which is high-dimensional data stored in a database, as input data to an input layer, and performs encoding of the high-dimensional descriptor into a latent variable, which is low-dimensional data, through an encoder. Next, a decoder of the CVAE 1000 decodes the low-dimensional latent variable, and finally, a descriptor, which is high-dimensional data corresponding to a new chemical structure, is output from an output layer.

[0135] Specifically, the CVAE 1000 may encode the high-dimensional descriptor x for each of the existing chemical structures into a low-dimensional latent variable l. The latent variable l may include a two-dimensional descriptor z and core structure information t obtained by reducing the dimension of the high-dimensional descriptor x. However, the data included in the latent variable l is not limited thereto.

[0136] An example of the structure corresponding to the core structure information t is shown in Table 1 below.

[0137] Table 1

[0138]

[0139]

[0140] The latent variable l can be plotted as a latent map 1030. The two-dimensional descriptor z included in the latent variable l corresponds to the (x, y) coordinate value of the latent map 1030, and the core structure information t included in the latent variable l is distinguished by color in the latent map 1030 (where Fig.10 The color of 1030 of the potential map is not shown).

[0141] According to an embodiment, the core structure information t may be displayed in eight different colors in the potential map 1030. For example, a latent variable l having a core structure t=0 may be displayed in blue in the potential map 1030, and a latent variable l having a core structure t=1 may be displayed in sky blue in the potential map 1030.

[0142] When using the potential graph 1030, the similarities between the descriptors can be visually identified. In the potential graph 1030, it can be confirmed that the descriptors are clustered (aggregated) according to the structural form. Therefore, it is relatively easy to generate new chemical structures including specific structural forms. Fig.10 From the potential graph 1030 , it can be confirmed that descriptors are clustered according to the same color, ie, the same structural form.

[0143] The CVAE 1000 may select a region 1040 to which latent variables are not mapped from the latent map 1030. When the latent variables 1040 of the selected region are decoded, the latent variables may be converted into high-dimensional descriptors. The CVAE 1000 may generate a chemical structure for the converted high-dimensional descriptors, thereby generating (creating) a new chemical structure.

[0144] That is, the CVAE 1000 can generate a new chemical structure not stored in the database by decoding a predetermined latent variable of the region 1040 of the potential graph 1030 to which the latent variables are not mapped. In addition, a chemical structure including a specific structure can be generated by decoding a latent variable of a region of the potential graph 1030 adjacent to a region in which the latent variables of the specific structure are clustered.

[0145] In addition, the neural network device can generate a property value of the new chemical structure generated by the CVAE 100 by using the DNN 1010 and generate its structural feature value by using the RNN 1020. The neural network device can also evaluate the properties and structural characteristics of the new chemical structure. When the target property value and the target structural feature value are not satisfied as a result of the evaluation, the neural network device generates a next generation chemical structure by applying a genetic algorithm to the generated new chemical structure.

[0146] Fig.11 is a flowchart for describing a method of generating a chemical structure by using a DNN and a RNN according to an embodiment. Fig.11 The method for generating a chemical structure by a neural network device shown in FIG. is related to the embodiment described above with reference to the accompanying drawings. Therefore, although omitted in the following description, the contents described above with reference to the accompanying drawings can also be applied to Fig.11 The method shown in .

[0147] Reference Fig.11 , the neural network device may receive a target property value and a target structural feature value as input data in operation 1110 .

[0148] The target property value is a numerical value of a certain property possessed by the chemical structure to be ultimately generated by the neural network device. According to one embodiment, the target property value may be a refractive index value, elastic modulus, melting point, transmission wavelength, and emission wavelength. According to another embodiment, the target property value may be set to an increasing (+) direction or a decreasing (-) direction instead of a given numerical value.

[0149] The target structural feature value is a numerical value of the local structure that must be included in the chemical structure that will be finally generated by the neural network device. The local structure can be expressed numerically using descriptors such as ECFP and QSPR.

[0150] The neural network device may generate a chemical structure including a specific local structure and having improved properties based on the target property value and the target structural feature value.

[0151] In operation 1120, the neural network device may select a first generation descriptor based on the received input data. The neural network device may select a descriptor having a property value similar to the received target property value and a structural feature value similar to the received target structural feature value from among the descriptors stored in the memory as the first generation descriptor.

[0152] In operation 1130, the neural network device may generate a second-generation descriptor by applying a genetic algorithm to the first-generation descriptor. The neural network device may generate the second-generation descriptor by performing crossover and mutation operations on the first-generation descriptor.

[0153] In operation 1140 , the neural network device may evaluate the properties and structure of the second generation descriptors.

[0154] The neural network device can generate a property value of each second-generation descriptor as output data by inputting the second-generation descriptor to the DNN as input data and applying an encoding function and a prediction function thereto. The neural network device can evaluate the property of the second-generation descriptor by comparing the generated property value with a target property value.

[0155] In addition, the neural network device can generate a structural feature value of each second-generation descriptor as output data by inputting the second-generation descriptor to the RNN as input data and applying an encoding function and a decoding function thereto. The neural network device can evaluate the structure of the second-generation descriptor by comparing the generated structural feature value with a target structural feature value.

[0156] According to an embodiment, the neural network device may determine whether the structural feature value includes a target structural feature value by comparing a SMILES code corresponding to the structural feature value with a SMARTS code corresponding to the target structural feature value using the RDkit library. When the structural feature value generated by the RNN 910 includes the target structural feature value, that is, when the generated final chemical structure includes the local structure, the neural network device may select the second generation descriptor input to the RNN 910 as a candidate descriptor.

[0157] In operation 1150 , the neural network apparatus may select a candidate descriptor having a target property value and a target structure feature value from the second generation descriptors based on the evaluation result.

[0158] Meanwhile, when the property value and / or structural feature value of the second generation descriptor cannot reach the target value, the neural network device may generate a third generation descriptor by applying a genetic algorithm to the second generation descriptor. The neural network device may evaluate the property of the third generation descriptor by using a DNN and the structure of the third generation descriptor by using an RNN. That is, the neural network device may develop the descriptor by using a genetic algorithm until a descriptor that satisfies the target property value and the target structural feature value is generated.

[0159] In operation 1160 , the neural network device may generate a chemical structure for the selected candidate descriptor.

[0160] The neural network device can generate a structural feature value of each candidate descriptor by inputting the candidate descriptor to the RNN as input data and applying an encoding function and a decoding function thereto. The neural network device can generate a chemical structure by converting the generated structural feature value into a SMILES code, an InChi code, etc.

[0161] Fig.12 is a flow chart for describing a method of generating a chemical structure by using CVAE according to an embodiment. Fig.12 The method for generating a chemical structure by a neural network device shown in FIG. 1 is related to the embodiment described above with reference to the accompanying drawings. Therefore, although omitted in the following description, the contents described above with reference to the accompanying drawings can also be applied to Fig.12 The method shown in .

[0162] Reference Fig.12 , the neural network device may convert the first descriptor into a low-dimensional latent variable in operation 1210.

[0163] The memory stores the structures (structural feature values) and properties (property values) that match each other as a group. The structural feature values ​​are stored in the memory in the form of descriptors, and the descriptors stored in the memory may be high-dimensional data.

[0164] The neural network device may convert the descriptors of the chemical structure stored in the memory (i.e., first descriptors) into low-dimensional latent variables, respectively. According to an embodiment, after receiving the target property value and / or the target structure feature value as input data, the neural network device may select some of the descriptors of the chemical structure stored in the memory as the first descriptors based on the received input. The neural network device may convert the selected first descriptors into low-dimensional latent variables.

[0165] The neural network device can encode the high-dimensional descriptor into a low-dimensional latent variable by using CVAE. The latent variable may include a two-dimensional descriptor and a core structure obtained by reducing the dimension of the high-dimensional descriptor.

[0166] In operation 1220, the neural network device may plot the latent variables into a latent graph.

[0167] The two-dimensional descriptor included in the latent variable may correspond to the (x, y) coordinate value of the latent map, and the core structure included in the latent variable may be expressed as a color in the latent map.

[0168] In operation 1230 , the neural network apparatus may select a region to which latent variables are not mapped from the latent map.

[0169] The neural network device may input predetermined values ​​included in a region to which latent variables of the potential graph are not mapped to a decoder and decode the predetermined values, thereby generating a new chemical structure not stored in a database.

[0170] Since it can be confirmed that descriptors are clustered according to structural forms in the potential map, it is relatively easy to generate a new chemical structure therefrom. A chemical structure including a specific structure can be generated by decoding a potential variable in a region adjacent to a region in which potential variables of a specific structure are clustered among potential variables included in a region to which potential variables of an existing chemical structure are not mapped.

[0171] In operation 1240 , the neural network device may convert the latent variables of the selected region into a high-dimensional second descriptor and generate a chemical structure for the second descriptor.

[0172] The neural network device may convert the latent variable into a high-dimensional descriptor by decoding the latent variable of the selected region using CVAE. In addition, the neural network device may generate a new chemical structure by generating a chemical structure for the converted high-dimensional descriptor using CVAE.

[0173] In addition, the neural network device may evaluate the properties and structures of each converted high-dimensional descriptor. When, as a result of the evaluation, the target property value and the target structure feature value are not satisfied, the neural network device may generate a next-generation chemical structure by applying a genetic algorithm to the converted high-dimensional descriptor. The description thereof is the same as that of operations 1110 to 1160 and will not be repeated here.

[0174] In addition, the foregoing embodiments may include computer-executable instructions, such as program modules executed by a computer in the form of a recording medium. Computer-readable media may be any recording medium that can be accessed by a computer, and may include volatile and non-volatile media and removable and non-removable media. In addition, the computer-readable media may include computer storage media and communication media. The computer storage media includes volatile and non-volatile and removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. The communication medium includes computer-readable instructions, data structures, program modules, or other data modulated by a data signal or other transmission mechanism and may include any transmission medium.

[0175] In addition, throughout the specification, the term "unit" may be a hardware component such as a processor or a circuit and / or a software component executed by a hardware component such as a processor.

[0176] The above description of the present disclosure is provided for illustrative purposes, and those skilled in the art will appreciate that various changes and modifications may be made without changing the technical concept and essential features of the present disclosure. Therefore, it is clear that the above illustrative embodiments are illustrative and do not limit the present disclosure in all aspects. For example, each component described as a single type can be implemented in a distributed manner. Similarly, the components described as distributed can be implemented in a combined manner.

[0177] It should be understood that the embodiments described herein should be considered in a descriptive sense only and not for purposes of limitation. Descriptions of features or aspects in each embodiment should typically be considered as available for other similar features or aspects in other embodiments.

[0178] Although one or more embodiments have been described with reference to the drawings, it will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope as defined by the following claims.

Claims

1. A method for generating a chemical structure, the method being performed by a neural network device and comprising: receiving a target property value and a target structural feature value as separate inputs; Selecting a first generation descriptor based on the target property value and the target structural feature value; generating a second generation descriptor by applying a genetic algorithm to the first generation descriptor; evaluating a property of each second generation descriptor by using a first neural network and evaluating a structure of each second generation descriptor by using a second neural network; selecting a candidate descriptor having the target property value and the target structural feature value from the second generation descriptors based on the evaluation result; and A chemical structure is generated for the selected candidate descriptors by using the second neural network.

2. The method of claim 1, wherein the selecting of the candidate descriptors further comprises: When the number of the selected candidate descriptors is less than a predetermined value, changing a transformation index of the genetic algorithm; and A second generation descriptor is regenerated by applying a genetic algorithm to the first generation descriptor based on the changed transformation index.

3. The method of claim 1, wherein the evaluation of the property comprises: inputting the second generation descriptors into the first neural network and generating a property value of each second generation descriptor as an output; and evaluating the property of each second generation descriptor by comparing the generated property value with the target property value, The first neural network is a deep neural network (DNN).

4. The method of claim 1, wherein the evaluation of the structure comprises: inputting the second generation descriptors into the second neural network and generating structural feature values ​​of each second generation descriptor as output; and evaluating the structure of each second generation descriptor by determining whether the generated structural feature value includes the target structural feature value, The second neural network is a recurrent neural network (RNN).

5. A neural network device configured to generate a chemical structure, the neural network device comprising: a user interface configured to receive a target property value and a target structural feature value as separate inputs; a memory storing at least one program; and a processor configured to drive the neural network device by executing the at least one program, wherein the processor selects a first generation descriptor based on the target property value and the target structural feature value, generating a second generation descriptor by applying a genetic algorithm to the first generation descriptor, evaluating the property of each second-generation descriptor by using a first neural network, evaluating the structure of each second-generation descriptor by using a second neural network, and selecting a candidate descriptor having the target property value and the target structure feature value from the second-generation descriptors based on the evaluation results, and A chemical structure is generated for the selected candidate descriptors by using the second neural network.

6. The neural network device of claim 5, wherein when the number of selected candidate descriptors is less than a predetermined value, the processor changes a transformation index of the genetic algorithm; and A second generation descriptor is regenerated by applying a genetic algorithm to the first generation descriptor based on the changed transformation index.

7. The neural network device of claim 5, wherein the processor inputs the second generation descriptors to the first neural network and generates a property value of each second generation descriptor as an output; and evaluating the property of each second generation descriptor by comparing the generated property value with the target property value, The first neural network is a deep neural network (DNN).

8. The neural network device of claim 5, wherein the processor inputs the second generation descriptors to the second neural network and generates a structural feature value of each second generation descriptor as an output; and evaluating the structure of each second generation descriptor by determining whether the generated structural feature value includes the target structural feature value, The second neural network is a recurrent neural network (RNN).

9. A method for generating chemical structures by using a conditional variational autoencoder (CVAE) using a neural network device, the method comprising: receiving a target property value and a target structural feature value as separate inputs; Converting the first descriptor into a low-dimensional latent variable; plotting the latent variables into a latent graph; selecting, from the latent map, a region to which the latent variable transformed by the first descriptor is not mapped; converting the latent variable of the selected region into a high-dimensional second descriptor; evaluating a property of each second descriptor by using a first neural network and evaluating a structure of each second descriptor by using a second neural network; and A chemical structure is generated for the second descriptor.

10. The method of claim 9, wherein the low-dimensional latent variables include core structure information and low-dimensional descriptors.

11. The method of claim 9, wherein a candidate descriptor having the target property value and the target structure feature value is selected from the second descriptor based on a result of the evaluation; and A chemical structure is generated for the selected candidate descriptors by using the second neural network.

12. The method of claim 9, further comprising: generating a next generation descriptor by applying a genetic algorithm to the second descriptor; evaluating a property of each next generation descriptor by using a first neural network, and evaluating a structure of each next generation descriptor by using a second neural network; selecting a candidate descriptor having the target property value from the next generation descriptors based on the evaluation result; and A chemical structure is generated for the selected candidate descriptors by using the second neural network.

13. The method of claim 11, wherein the evaluation of the property comprises: inputting the second descriptors into the first neural network and generating a property value of each second descriptor as an output; as well as evaluating the property of each second descriptor by comparing the generated property value with the target property value, The first neural network is a deep neural network (DNN).

14. The method of claim 11, wherein the evaluation of the structure comprises: inputting the second descriptors into the second neural network and generating a chemical structure value for each second descriptor as an output; as well as evaluating the structure of the second descriptor by determining whether the generated structural feature values ​​include the target structural feature values, The second neural network is a recurrent neural network (RNN).

15. A neural network device configured to generate a chemical structure by using a conditional variational autoencoder (CVAE), the neural network device comprising: a memory storing at least one program; and a processor configured to drive the neural network device by executing the at least one program, wherein the processor receives a target property value and a target structural feature value as separate inputs; Converting the first descriptor into a low-dimensional latent variable; plotting the latent variables into a latent graph; selecting, from the latent map, a region to which the latent variable transformed by the first descriptor is not mapped; converting the latent variable of the selected region into a high-dimensional second descriptor; evaluating a property of each second descriptor by using a first neural network and evaluating a structure of each second descriptor by using a second neural network; and A chemical structure is generated for the second descriptor.

16. The neural network device of claim 15, wherein the low-dimensional latent variables include core structure information and low-dimensional descriptors.

17. The neural network device of claim 15, further comprising a user interface configured to receive the target property value and the target structural feature value as input, wherein the processor further selects a candidate descriptor having the target property value and the target structural feature value from the second descriptor based on a result of the evaluation, and A chemical structure is generated for the selected candidate descriptors by using the second neural network.

18. The neural network device of claim 15, further comprising a user interface configured to receive the target property value and the target structural feature value as input, wherein the processor further generates a next generation descriptor by applying a genetic algorithm to the second descriptor; evaluating a property of each next generation descriptor by using the first neural network, and evaluating a structure of each next generation descriptor by using a second neural network; selecting a candidate descriptor having the target property value and the target structural feature value from the next generation descriptor based on the evaluation result; and A chemical structure is generated for the selected candidate descriptors by using the second neural network.

19. The neural network device of claim 17, wherein the processor inputs the second descriptor into a recurrent neural network and generates a chemical structure as output for each high-dimensional descriptor, and The structure of the second descriptor is evaluated by determining whether the generated structure feature values ​​include the target structure feature values.

20. The neural network device of claim 17, wherein the processor inputs the second descriptors into a deep neural network and generates a property value of each second descriptor as an output, and The property of each second descriptor is evaluated by comparing the generated property value with the target property value.

21. A computer-readable recording medium comprising a program which, when executed by a computer, implements the method according to any one of claims 1 to 4 and 9 to 14.

Images