Apparatus and method for generating new chemical structures using neural networks

Abstract

To provide a method and device for generating a new chemical structure using a neural network.SOLUTION: A neural network device of the present invention for generating a new chemical structure is configured to: receive a structure input of a chemical structure; generate a negative attention vector that indicates a respective probability of presence of each of a plurality of blacklists in the structure input; generate a structure expression by encoding the structure input; generate a final reverse blacklist vector that does not include the plurality of blacklists on the basis of the negative attention vector and the structure expression; and generate a new chemical structure by decoding the final reverse blacklist vector.SELECTED DRAWING: Figure 8

Description

[Technical field]

[0001] The present invention relates to an apparatus and method for generating new chemical structures using a neural network. [Background technology]

[0002] A neural network is a computational architecture that models the biological brain. With the development of neural network technology, many electronic systems use neural networks to analyze input data and extract useful information.

[0003] Recently, research is being actively conducted on selecting chemical structures to be used in material development by evaluating the physical properties of chemical structures using neural network technology. In the conventional method of designing materials using neural network technology, a new chemical structure is first obtained, and then a post-mortem examination is performed to determine whether the new chemical structure includes a predetermined blacklist, followed by filtering.

[0004] However, most new chemical structures generated through a neural network model include a blacklist, making it difficult to obtain new chemical structures with satisfactory properties. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] International Publication No. 2017 / 094899 Summary of the Invention [Problem to be solved by the invention]

[0006] The present invention has been made in consideration of the above-mentioned problems in the conventional art, and an object of the present invention is to provide an apparatus and method for generating chemical structures using a neural network, and to provide a computer-readable recording medium having recorded thereon a program for causing a computer to execute the method. [Means for solving the problem]

[0007] In order to achieve the above-mentioned object, a method for generating a new chemical structure using a neural network device according to one aspect of the present invention includes a step of receiving a structural input of a chemical structure, a step of generating a negative attention vector based on the structural input, the negative attention vector representing each probability of the presence of each of a plurality of blacklists in the structural input, a step of encoding the structural input to generate a structural representation, a step of generating a final inverse blacklist vector that does not include the plurality of blacklists based on the negative attention vector and the structural representation, and a step of decoding the final inverse blacklist vector to generate a new chemical structure.

[0008] The step of generating the final reverse blacklist vector may include a step of calculating a reverse negative attention vector using the negative attention vector, and a step of generating a final reverse blacklist vector that does not include the multiple blacklists through an operation between the reverse negative attention vector and the structural representation. The step of generating the final inverse blacklist vector includes a step of generating a final blacklist vector including the multiple blacklists through an operation between the negative attention vector and the structural representation during a learning process of a chemical structure generation model, and the step of generating the new chemical structure may further include a step of training the chemical structure generation model by feeding back prediction results of the new chemical structure and the blacklist corresponding to the new chemical structure. The step of training the chemical structure generation model may further include calculating a prediction result of the blacklist using a non-negative parameter. The receiving step may include a step of receiving, as the structure input, a chemical structure that has been confirmed not to include at least a portion of the plurality of blacklists during a learning process of a chemical structure generation model. The method may further include a step of selecting a portion of the plurality of blacklists, wherein generating the negative attention vector includes generating a negative attention vector based on the structure input, the negative attention vector representing a respective probability of each of the selected blacklists being present in the structure input, and generating the final reverse blacklist vector may include generating a final reverse blacklist vector that does not include the selected blacklist based on the negative attention vector and the structure representation. Generating the final reverse blacklist vector may include generating the final reverse blacklist vector via an elementwise multiplication operation of the reverse negative attention vector and the structural representation. The method may further include generating a positive attention vector based on the structure input, the positive attention vector representing a respective probability of the presence of each of a plurality of whitelists in the structure input; generating a final whitelist vector based on the positive attention vector and the structure representation, the final whitelist vector including the plurality of whitelists; and decoding the final whitelist vector to generate a new chemical structure. The method may further include a step of receiving, as the structure input, a chemical structure that has been confirmed to include at least a portion of the plurality of whitelists during a training process of a chemical structure generation model. The method may further include a step of receiving as the structure input a chemical structure that has been confirmed to not include at least a portion of the multiple blacklists and to include at least a portion of the multiple whitelists during a learning process of a chemical structure generation model.

[0009] In order to achieve the above-mentioned object, one aspect of the present invention provides a neural network device for generating new chemical structures, comprising a memory in which at least one program is stored, and a processor for driving a neural network, wherein the processor is configured to receive a structural input of a chemical structure by executing the at least one program, generate a negative attention vector representing each probability of the presence of each of a plurality of blacklists in the structural input based on the structural input, encode the structural input to generate a structural representation, generate a final inverse blacklist vector representing a portion that does not correspond to the plurality of blacklists based on the negative attention vector and the structural representation, and generate a new chemical structure based on the final inverse blacklist vector.

[0010] In order to achieve the above object, a computer-readable recording medium having a program recorded thereon according to one aspect of the present invention causes a computer to execute the above method. Effect of the Invention

[0011] According to the present invention, a final inverse blacklist vector that does not include a blacklist is generated using the inverse negative attention vector, and a new chemical structure that does not include a blacklist that is disadvantageous for synthesis can be generated by decoding the final inverse blacklist vector. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a block diagram showing a hardware configuration of a neural network device according to an embodiment. [Diagram 2]FIG. 1 is a diagram illustrating operations performed in a deep neural network (DNN) according to one embodiment. [Diagram 3] FIG. 1 is a diagram illustrating operations performed in a recurrent neural network (RNN) according to one embodiment. [Figure 4] FIG. 2 is a diagram for explaining operations performed in a conditional variational autoencoder (CVAE) according to one embodiment. [Diagram 5] FIG. 1 is a conceptual diagram illustrating a neural network system for generating chemical structures according to one embodiment. [Figure 6] FIG. 13 is a diagram for explaining an example of a process for generating a new chemical structure using a final blacklist vector according to one embodiment. [Figure 7] FIG. 2 illustrates an example of a blacklist according to one embodiment. [Figure 8] FIG. 13 is a diagram illustrating another example of a process for generating a new chemical structure using a final reverse blacklist vector according to an embodiment. [Figure 9] FIG. 13 is a diagram illustrating a process of generating a new chemical structure using a final whitelist vector according to one embodiment. [Figure 10] 1 is a flow chart illustrating a method for generating new chemical structures in a neural network device according to one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] The appearances of phrases such as "in some embodiments" or "in one embodiment" in various places in the specification do not necessarily both refer to the same embodiment.

[0014] Some embodiments of the present invention are illustrated by functional block configurations and various processing steps. Some or all of such functional blocks may be realized by various numbers of hardware and / or software configurations performing specific functions. For example, the functional blocks of the present invention may be realized by one or more microprocessors or by circuit configurations for a given function. For example, the functional blocks of the present invention may be realized by various programming or scripting languages. The functional blocks may be realized by algorithms executed by one or more processors. The present invention may employ conventional techniques for electronic configuration, signal processing, and / or data processing. Terms such as "mechanism," "element," "means," and "configuration" are generic and are not limited to mechanical and physical configurations.

[0015] In addition, the connecting lines or connecting members between components shown in the drawings are illustrative of functional connections and / or physical or circuit connections, and in an actual device, connections between components are shown by various functional connections, physical connections, or circuit connections that can be replaced or added.

[0016] Meanwhile, with regard to the terminology used in this specification, a descriptor, which is data used in a neural network system, is an index value used to express the characteristics of a substance, and is a value obtained by performing a relatively simple arithmetic process on a given substance. In one embodiment, the descriptor includes a molecular structure fingerprint (e.g., Morgan Fingerprint: ECFP (Extended Connectivity Fingerprint)) that indicates whether or not a specific substructure is included, or a quantitative structure-property relationship (QSPR) descriptor that is composed of a value that is instantly calculated, such as a molecular weight or the number of substructures (e.g., rings) included in a molecular structure.

[0017] Furthermore, a property means a characteristic of a material, and is a real value measured through an experiment or calculated through a simulation. For example, if the material is a display material, it is a transmission wavelength or an emission wavelength for light, and if the material is a battery material, it is a voltage. Unlike the case of an expression, calculation of a property requires a complex simulation, which takes a long time.

[0018] Additionally, structure refers to the atomic level structure of a substance. Since it is required to express the structure at the atomic level in order to derive physical properties through first principles calculations, the structure of a substance must be derived at the atomic level in order to generate a new chemical structure. A structure is a structural formula based on the bonds between atoms, and is a simple string (one-dimensional) format. String formats for expressing structures include SMILES (simplified molecular-input line-entry system) code, SMARTS (SMILES arbitrary target specification) code, and InChi (international chemical identifier) ​​code.

[0019] In addition, factors are elements that define the relationship between expressions, physical properties, and structures. Factors are determined by machine learning based on the expressions-physical properties-structural formula relationships stored in the database. This allows us to know how factors are connected to expressions, physical properties, and structural formulas.

[0020] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, specific examples of the embodiments of the present invention will be described in detail with reference to the drawings.

[0021] FIG. 1 is a block diagram showing a hardware configuration of a neural network device according to an embodiment.

[0022] The neural network device 100 is embodied by various devices such as a personal computer (PC), a server device, a mobile device, and an embedded device, and specific examples thereof include, but are not limited to, smartphones, tablet devices, augmented reality (AR) devices, internet of things (IoT) devices, autonomous vehicles, robotics, medical devices, etc., which perform voice recognition, image recognition, image classification, etc., using a neural network. Furthermore, the neural network device 100 corresponds to a dedicated hardware accelerator (HW accelerator) mounted on the above-mentioned devices, and the neural network device 100 is a hardware accelerator such as a neural processing unit (NPU), a tensor processing unit (TPU), or a neural engine, which is a dedicated module for driving a neural network, but is not limited to these.

[0023] 1, a neural network device 100 includes a processor 110 and a memory 120. Only components related to this embodiment are shown in the neural network device 100 shown in Fig. 1. Therefore, it will be obvious to a person skilled in the art that the neural network device 100 may further include other general-purpose components in addition to the components shown in Fig. 1.

[0024] The processor 110 serves to control the overall functions for executing the neural network device 100. For example, the processor 110 controls the overall neural network device 100 by executing a program stored in the memory 120 in the neural network device 100. The processor 110 may be realized by a central processing unit (CPU), a graphics processing unit (GPU), an application processor (AP), etc. provided in the neural network device 100, but is not limited thereto.

[0025] The memory 120 is hardware that stores various data processed within the neural network device 100. For example, the memory 120 stores data that has been processed by the neural network device 100 and data to be processed. The memory 120 also stores applications, drivers, etc. that are run by the neural network device 100. The memory 120 includes random access memory (RAM) such as dynamic random access memory (DRAM) or static random access memory (SRAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), compact disc read only memory (CD-ROM), Blu-ray or other optical disc storage, hard disk drive (HDD), solid static disk drive (SSD), or flash memory.

[0026] Meanwhile, in the memory 120, the structural feature values ​​and the physical property values ​​are matched and stored as one set, and the neural network device 100 reads the structural feature values ​​and the physical property values ​​from the memory 120 or writes the structural feature values ​​and the physical property values ​​to the memory 120. The physical property values ​​indicate properties possessed by a substance and are real values ​​measured through experiments or calculated through simulations. In addition, the structural feature values ​​are index values ​​used to express the structure of a substance, and are molecular structure fingerprints (e.g., Morgan Fingerprint: ECFP) that indicate whether a specific substructure is included or not.

[0027] The processor 110 runs a deep neural network (DNN), a recurrent neural network (RNN), and a conditional variational autoencoder (CVAE).

[0028] The processor 110 trains a deep neural network (DNN) using the representations and physical properties (physical property values), and in the process, determines factors that define the relationship between the representations and the physical properties. The processor 110 drives the trained deep neural network (DNN) to perform calculations using the representations as input data in the input layer, and generates physical property values ​​as output data based on the results of the calculations.

[0029] The processor 110 trains a recurrent neural network (RNN) using the expressions and structures (structural feature values), and in the process, determines factors that define the relationship between the expressions and structures. The processor 110 drives the trained recurrent neural network (RNN) to perform calculations in the input layer using the expressions or factors as input data, and generates structural feature values ​​as output data based on the results of the calculations.

[0030] The processor 110 uses the CVAE to generate a new chemical structure that does not exist in the database. Specifically, in the input layer of the CVAE, a representation, which is high-dimensional data stored in the database, is used as input data, and an encoder performs encoding to convert the high-dimensional representation into a latent variable having a lower dimension. Then, a decoder of the CVAE decodes the low-dimensional latent variable, and finally, a representation, which is high-dimensional data corresponding to the new chemical structure, is output in the output layer.

[0031] Meanwhile, the neural network device 100 further includes a user interface (not shown). The user interface means a means for inputting data for controlling the neural network device 100. For example, the user interface may include, but is not limited to, a key pad, a dome switch, a touch pad (contact type capacitance type, pressure type resistive film type, infrared sensing type, surface ultrasonic conduction type, integral type tension measurement type, piezoelectric effect type, etc.), a jog wheel, a jog switch, etc.

[0032] In the following description of the present embodiment, a method for generating a chemical structure and evaluating the generated chemical structure using the neural network device 100 will be described in detail. The method described below is performed by the processor 110 and memory 120 of the neural network device 100.

[0033] FIG. 2 is a diagram illustrating operations performed in a deep neural network (DNN) according to one embodiment.

[0034] Referring to FIG. 2, a deep neural network (DNN) 20 has a structure including an input layer, a hidden layer, and an output layer, and receives input data (e.g., I 1 and I 2 ) and outputs the output data (e.g., O 1 and O 2 )

[0035] For example, as shown in FIG. 2, the deep neural network (DNN) 20 includes an input layer Layer1, two hidden layers Layer2 and Layer3, and an output layer Layer4. Since the deep neural network (DNN) 20 includes more layers that process useful information, the deep neural network (DNN) 20 can process a more complex data set than a neural network with a single layer. On the other hand, the deep neural network (DNN) 20 is illustrated as including four layers, but this is merely an example, and the deep neural network (DNN) 20 may include fewer or more layers or fewer or more channels. That is, the deep neural network (DNN) 20 may include layers with various structures different from those shown in FIG. 2.

[0036] Each of the layers included in the deep neural network (DNN) 20 includes a plurality of channels. The channels correspond to a plurality of artificial nodes known as neurons, processing elements (PEs), units, or similar terms. For example, as shown in FIG. 2, Layer 1 includes two channels (nodes), and Layer 2 and Layer 3 each include three channels. However, this is merely an example, and each of the layers included in the deep neural network (DNN) 20 includes various numbers of channels (nodes).

[0037] The channels included in each layer of the deep neural network (DNN) 20 are connected to each other to process data. For example, one channel receives data from another channel, performs a calculation, and outputs the calculation result to yet another channel.

[0038] The inputs and outputs of each channel are called input activations and output activations, respectively. That is, activations are parameters that correspond to the output of one channel and the input of the channel in the next layer. Meanwhile, each channel determines its own activation based on the activations and weights received from the channel in the previous layer. Weights are parameters used to calculate the output activations in each channel and are values ​​assigned to the connection relationships between channels.

[0039] Each of the channels is processed by a computational unit or processing element that receives inputs and outputs output activations, and the inputs and outputs of each of the channels are mapped, e.g., where σ is the activation function, TIFF0007665322000001.tif9128 is the weight from the kth channel in the (i-1)th layer to the jth channel in the ith layer, TIFF0007665322000002.tif9128 is the bias of the jth channel in the ith layer, If TIFF0007665322000003.tif9128 is the activation of the jth channel of the ith layer, then the activation TIFF0007665322000004.tif9128 is calculated using the following formula 1.

[0040]

number

[0041] As shown in Figure 2, the activation of the first channel CH1 in the second layer Layer2 is It is expressed as TIFF0007665322000006.tif9128. Also, TIFF0007665322000007.tif9128 is calculated as follows according to formula 1: TIFF0007665322000008.tif18130. However, the above formula 1 is an example of the explanation of the activation and weights used for data processing in the deep neural network (DNN) 20, and is not limited thereto. The activation is a value obtained by applying an activation function to the sum of the activations received from the previous layer and passing the result through a rectified linear unit (ReLU).

[0042] In one embodiment, the deep neural network (DNN) 20 determines factors that define the relationship between the representations and physical properties through learning using the representations and physical property values. That is, in Layer 1 to Layer 4 that constitute the deep neural network (DNN) 20, the representations become Layer 1, which is an input layer, the physical property values ​​become Layer 4, which is an output layer, and the factors become at least one hidden layer, Layer 2 and / or Layer 3.

[0043] In the deep neural network (DNN) 20, in the input layer, a calculation is performed using the representation as input data, and physical property values ​​are generated as output data based on the results of the calculation.

[0044] FIG. 3 is a diagram illustrating the operations performed in a recurrent neural network (RNN) according to one embodiment.

[0045] Hereinafter, explanations that overlap with those in FIG. 2 will be omitted for convenience.

[0046] The recurrent neural network (RNN) 30 is a neural network for learning data that changes over time, such as time-series data, and is configured by connecting a network to a reference time point t and a next time point t+1. That is, the recurrent neural network (RNN) 30 is a neural network that takes into account the time aspect, capable of effectively learning patterns from data that is input sequentially or in which a sequence of features is given as input, by modifying a model to enable recursive input in a hidden layer of the neural network.

[0047] 3, there is shown a node s constituting a hidden layer of a recurrent neural network (RNN) 30. The node s performs a calculation based on input data x and generates output data o. The recurrent neural network (RNN) 30 applies the same task to every sequence, and the final output result of the node s is affected by the results of the previous calculation.

[0048] The recurrent neural network (RNN) 31 is an unfolding of the recurrent neural network (RNN) 30 having a loop. "Unfolding" the recurrent neural network (RNN) 30 means that the recurrent neural network (RNN) 30 is expressed for the entire sequence. In the recurrent neural network (RNN) 31, x t is the input value at time step t, and s t is the hidden state at time step t. t is expressed as the following Equation 2, where tanh or Relu is used as the function f, and s t-1 is generally initialized to 0. Also, in a recurrent neural network (RNN)31, o t is the output value at time step t.

[0049]

number

[0050] s t is the memory part of the network, and s t contains information about what happened in the previous time step, and the output value o t depends only on the memory for the current time step t.

[0051] On the other hand, the parameter values ​​for each layer are different from other existing neural network structures, and the recurrent neural network (RNN) 31 shares the parameter values ​​U, V, and W for all time steps. That is, in each step of the recurrent neural network (RNN) 31, only the input values ​​are different and approximately the same calculations are performed, so that the number of parameters that must be learned can be reduced.

[0052] In one embodiment, the recurrent neural network (RNN) 32 determines factors that define the relationship between the expressions and the structures through learning using the expressions and the structures (structural feature values). As previously described in FIG. 2, the factors are at least one hidden layer. In the input layer of the recurrent neural network (RNN) 32, the expressions or factors are used as input data to perform calculations, and structural feature values ​​are generated as output data based on the results of the calculations.

[0053] For example, if the character string indicating the structural feature value is "ABCDEFG," the input and output for each time period will be "ABC," "BCD," "CDE," "DEF," and "EFG." That is, each step of the recursive neural network (RNN) part uses the character string following the character string input at time t as the input at time (t+1).

[0054] The recurrent neural network (RNN) 32 performs learning so as to maximize the probability of generating a character string showing the structure s for the data (h, s) (i.e., factor and structure data). The probability p of generating a character string is expressed by the following Equation 3.

[0055]

number

[0056] When the configuration of the recurrent neural network (RNN) part is completed through learning (i.e., when the parameter values ​​of the recurrent neural network (RNN) model used for determining the structural factors are determined), the representation is encoded and the obtained factors are decoded. According to another embodiment, the decoding is sequentially performed by inputting the factor e(x) to obtain the first part (ABC) of the decoded string, inputting the obtained string "ABC" to obtain "BCD", and inputting the obtained string "BCD" to obtain the string "CDE".

[0057] 3, the first two characters of a string at a specific time t are the same as the last two characters of the string at time (t-1). Then, the strings at each time are aggregated and one string (ABCDEFG) is output.

[0058] FIG. 4 is a diagram illustrating the calculations performed in the CVAE according to one embodiment.

[0059] Referring to FIG. 4, the CVAE 40 has a structure including an input layer, an encoder, a decoder, and an output layer. In the input layer of the CVAE 40, high-dimensional data such as an image stored in a database is used as input data, and in the encoder, encoding is performed to convert the high-dimensional input data into a latent variable z of a lower dimension. In one embodiment, the latent variable z is generally 2-50 dimensional data according to a normal distribution with a mean μ and a variance σ. Then, in the decoder, a new image (high-dimensional data) not present in the database is output in the output layer by decoding the low-dimensional latent variable z.

[0060] For example, if an image of a human figure is used as input data, the latent variables would be the object shape, camera coordinates (view point), and light source information, etc., and if a number image is used as input data, the latent variables would be the line angle, aspect ratio, etc.

[0061] Meanwhile, the latent variable z is mapped to a latent map, and by inputting a predetermined value z' included in an area of ​​the latent map where the latent variable z is not mapped to the decoder and decoding it, new data that does not exist in the database is generated in the output layer.

[0062] In the input layer of the CVAE 40, the high-dimensional data, which is a representation stored in a database, is used as input data, and the encoder performs encoding to convert the high-dimensional representation into a latent variable having a lower dimension. Then, the decoder of the CVAE 40 decodes the low-dimensional latent variable, and finally, the output layer outputs a representation, which is high-dimensional data corresponding to a new chemical structure.

[0063] FIG. 5 is a conceptual diagram illustrating a neural network system for generating chemical structures according to one embodiment.

[0064] FIG. 5 illustrates a neural network system that generates chemical structures using a deep neural network (DNN) 500, a recurrent neural network (RNN) 510, and a CVAE 520.

[0065] Representations, which are data used in neural network systems, include quantitative structure-property relationship (QSPR) representations composed of instantly calculated values, such as molecular structure fingerprints (e.g., ECFP) that indicate whether or not a substructure is included. Properties refer to the characteristics of a substance and are real values ​​measured through experiments or calculated through simulations. Structure refers to the atomic-level structure of a substance. String formats that represent structures include SMILES code, SMARTS code, and InChi code. For example, the structural formula is represented by SMILES code as in Formula 4, or the structural formula is represented by SMARTS code as in Formula 5.

[0066]

number

[0067]

number

[0068] A factor is an element that specifies the relationship between a representation, a physical property, and a structure. A factor is at least one hidden layer, and when a factor includes multiple hidden layers, a factor that specifies the relationship between a representation and a physical property and a factor that specifies the relationship between a representation and a structure are determined for each hidden layer.

[0069] A latent map is a visual coordinate system in which two-dimensional and three-dimensional representations are mapped. Using a latent map, the similarity between representations can be visualized and confirmed.

[0070] In the deep neural network (DNN) 500, an input layer performs calculations using expression elements as input data, and generates physical properties (physical property values) as output data based on the results of the calculations. In the recurrent neural network (RNN) 510, an input layer performs calculations using expression elements or factors as input data, and generates structures (structural feature values) as output data based on the results of the calculations. SMILES codes and SMARTS codes are used as character string formats that indicate structural feature values.

[0071] In addition, the deep neural network (DNN) (50) and the recurrent neural network (RNN) (510) determine factors that are elements that define the relationship between the representation and the physical properties and the structure by performing learning based on the relationship between the physical properties (physical property values) and the structure (structural feature values) stored in the memory. In one embodiment, the factors are at least one hidden layer, and when the factors include multiple hidden layers, the factors that define the relationship between the representation and the physical properties and the factors that define the relationship between the representation and the structure are determined for each hidden layer.

[0072] The recurrent neural network (RNN) 510 performs a calculation using the factors determined in the deep neural network (DNN) 500 as input data, and generates a structural feature value as output data based on the calculation result.

[0073] In the CVAE520, the input layer uses the high-dimensional data, which is a representation stored in a database, as input data, and the encoder performs encoding to convert the high-dimensional representation into a latent variable having a lower dimension. Then, the decoder decodes the low-dimensional latent variable, and the output layer outputs a representation, which is high-dimensional data corresponding to a new chemical structure. When the representation output from the CVAE520 is used as input data to the deep neural network (DNN) 500, a physical property value can be generated, and when the output representation is used as input data to the recurrent neural network (RNN) 510, a structural feature value can be generated.

[0074] FIG. 6 is a diagram for explaining an example of a process for generating a new chemical structure using a final blacklist vector according to one embodiment.

[0075] The neural network device uses the blacklist in the learning and inference processes of the chemical structure generation model.

[0076] 6, the neural network device receives a structure input 610 of a chemical structure. In one embodiment, the neural network device receives a representation of a particular chemical structure as the structure input 610. The representation may include a molecular structure fingerprint (e.g., Morgan Fingerprint (ECFP)) or a quantitative structure-property relationship (QSPR) representation composed of immediately calculated values ​​such as the molecular weight or the number of substructures (e.g., rings) contained in the molecular structure. For example, the structure input 610 may be a 4000-dimensional vector composed of 0s or 1s.

[0077] The neural network device receives the blacklist. The blacklist refers to partial structures that are disadvantageous in synthesis and performance. Specifically, a chemical structure newly generated through the neural network device has a physical property value for a specific physical property (e.g., transmission wavelength, emission wavelength, etc.), but the neural network device aims to generate a new chemical structure having a physical property value that matches a target physical property value (e.g., "emission wavelength: λ=350 nm"). In this case, if the new chemical structure includes a partial structure that corresponds to the blacklist, the physical property value of the new chemical structure is highly likely to not match the target physical property value.

[0078] 7 is a diagram showing an example of a blacklist according to an embodiment. Referring to FIG. 7, for example, the blacklist may be "containing a ring structure with four or less corners", "containing six or more O atoms", "containing elements other than C, O, and N", and "containing a ring structure with seven or more corners". However, the blacklist is not limited to the above examples, and may be set in various ways depending on the desired composition and performance.

[0079] Based on the structure input 610, the neural network device generates a negative attention vector 620 that represents a respective probability of the presence of each of a plurality of blacklists in the structure input 610.

[0080] The size of the negative attention vector 620 is determined by the number of blacklists. Each element of the negative attention vector 620 has a value between 0 and 1, and each element of the negative attention vector 620 indicates the probability that the corresponding blacklist exists in the structure input 610.

[0081] For example, if the blacklist consists of four items as shown in Fig. 7, the negative attention vector 620 is a four-dimensional vector [0.9 0.1 0.2 0.1]. In this case, the structure input 610 indicates that the probability (0.9) of the blacklist "containing a ring structure with four or fewer corners" is high, and the probability of the remaining blacklists is low.

[0082] The neural network device also encodes the structural input 610 to generate the structural representation 630. In one embodiment, the neural network device uses a CVAE encoder to encode the structural input 610 into the structural representation 630. For example, the neural network device encodes the structural input 610, which is a 4000-dimensional vector, into the structural representation 630, which is a 128-dimensional vector.

[0083] The neural network device generates a final blacklist vector 640 including multiple blacklists through an operation between the negative attention vector 620 and the structural representation 630. In one embodiment, the neural network device generates the final blacklist vector 640 through an elementwise operation between the negative attention vector 620 and the structural representation 630.

[0084] The neural network device decodes the final blacklist vector 640 to generate a new chemical structure 650. In one embodiment, the neural network device utilizes a CVAE decoder to decode the final blacklist vector 640 into a new chemical structure 650. The new chemical structure 650 is represented by a representation.

[0085] Meanwhile, the chemical structure generation model of the neural network device is trained using the final blacklist vector 640. Specifically, the neural network device generates a new chemical structure 650 by decoding the final blacklist vector 640. The neural network device analyzes the new chemical structure 650 to determine whether the new chemical structure 650 includes a blacklist that is highly likely to exist in the structure input 610. That is, the neural network device calculates a blacklist prediction result 660 by analyzing the new chemical structure 650.

[0086] The chemical structure generation model of the neural network device is trained by receiving as feedback a new chemical structure 650 generated from the final blacklist vector 640 and the blacklist prediction result 660 corresponding to the new chemical structure 650.

[0087] In addition, the neural network device inputs the new chemical structure 650 into the recurrent neural network (RNN) and obtains the structural feature value 670 of the new chemical structure 650 as the output value of the recurrent neural network (RNN).

[0088] The structural feature value 670 is an index value used to represent the structure of a substance, and is a molecular structural fingerprint (eg, Morgan Fingerprint "ECFP").

[0089] In addition, the neural network device inputs the new chemical structure 650 into a deep neural network (DNN) and obtains physical property values ​​680 of the new chemical structure 650 as output values ​​of the deep neural network (DNN).

[0090] The physical property value 680 means a numerical value of a specific physical property possessed by the new chemical structure 650. For example, the physical property value 680 is a refractive index value, an elastic coefficient, a melting point, a transmission wavelength, and an emission wavelength.

[0091] Meanwhile, when calculating the blacklist prediction result 660 by analyzing the new chemical structure 650, the neural network device calculates the blacklist prediction result 660 corresponding to the new chemical structure 650 using non-negative parameters. Using non-negative parameters means that a restrictive condition is set so that the values ​​of all parameters used in the neural network device do not become negative. For example, if the value of a specific parameter becomes negative, the neural network device applies an exponential function to the specific parameter to convert it into a non-negative value.

[0092] The learning performance of the chemical structure generation model is improved by the chemical structure generation model receiving as feedback the prediction results of the blacklist calculated using non-negative parameters.

[0093] In addition, in the learning process of the chemical structure generation model, the neural network device receives only chemical structures confirmed to not include at least a part of the plurality of blacklists in the structure input 610. For example, if there are a total of 13 blacklists, the neural network device receives only chemical structures that do not include all of the 13 blacklists in the structure input 610. In this way, the chemical structure generation model receives and learns only specific chemical structures that do not include at least a part of the blacklists in the structure input 610, thereby improving the learning performance of the chemical structure generation model.

[0094] FIG. 8 is a diagram for explaining another example of a process for generating a new chemical structure using a final reverse blacklist vector according to an embodiment.

[0095] 8, the neural network machine receives a structure input of a chemical structure 810. In one embodiment, the neural network machine receives a representation of a particular chemical structure at the structure input 810.

[0096] The neural network device receives the blacklist. Referring to FIG. 7, an example of the blacklist is shown. For example, the blacklist may be "containing a ring structure with four or fewer corners", "containing six or more O's", "containing elements other than C, O, and N", and "containing a ring structure with seven or more corners". However, the blacklist is not limited to the above examples, and may be set in various ways depending on the desired composition and performance.

[0097] Based on the structure input 810, the neural network device generates a negative attention vector 820 that represents a respective probability of the presence of each of a plurality of blacklists in the structure input 810.

[0098] The size of the negative attention vector 820 is determined by the number of blacklists. Each element of the negative attention vector 820 has a value between 0 and 1, and each element of the negative attention vector 820 indicates the probability that the corresponding blacklist exists in the structure input 810.

[0099] For example, if the blacklist consists of four items as shown in Figure 7, the negative attention vector 820 is a four-dimensional vector [0.9 0.1 0.2 0.1]. In this case, it can be seen that the probability (0.9) that the structure input 810 corresponds to the blacklist "including ring structures with four or fewer corners" is high, and the probability that it corresponds to the remaining blacklists is low.

[0100] The neural network device uses the negative attention vector 820 to calculate a reverse negative attention vector 860. In one embodiment, the neural network device calculates the reverse negative attention vector 860 as the difference between a basis vector and the negative attention vector 820. Here, the basis vector is a vector whose size is the same as that of the negative attention vector 820 and whose elements all have a value of 1.

[0101] For example, if the negative attention vector 820 is [0.9 0.1 0.2 0.1], the inverse negative attention vector 860 is [0.1 0.9 0.8 0.9]. That is, it means that there is a high possibility that the blacklist corresponding to the element whose value is close to 1 in the inverse negative attention vector 860 does not exist in the structure input 810.

[0102] The neural network device also encodes the structural input 810 to generate the structural representation 830. In one embodiment, the neural network device encodes the structural input 810 into the structural representation 830 using a CVAE encoder.

[0103] The neural network device generates a final inverse blacklist vector 840 that does not include multiple blacklists through an operation of the inverse negative attention vector 860 and the structural representation 830. In one embodiment, the neural network device generates the final inverse blacklist vector 840 through an element-wise operation of the inverse negative attention vector 860 and the structural representation 830.

[0104] The neural network device decodes the final inverse blacklist vector 840 to generate a new chemical structure 850. In one embodiment, the neural network device utilizes a CVAE decoder to decode the final inverse blacklist vector 840 into a new chemical structure 850. The new chemical structure 850 is represented by a representation.

[0105] As described above, in the reverse negative attention vector 860, the value of a specific element being close to 1 means that there is a high possibility that the blacklist corresponding to the specific element does not exist in the structure input 810. In other words, there is a high possibility that the new chemical structure 850 generated based on the reverse negative attention vector 860 does not include the blacklist contained in the structure input 810.

[0106] The blacklist represents a partial structure that is detrimental to synthesis and performance, but if the new chemical structure 850 does not include the blacklist, there is a very high probability that the physical property value 880 of the new chemical structure 850 will match the target physical property value.

[0107] The neural network device inputs a new chemical structure 850 into a recurrent neural network (RNN) and obtains a structural feature value 870 of the new chemical structure 850 as an output value of the recurrent neural network (RNN).

[0108] In addition, the neural network device inputs the new chemical structure 850 into a deep neural network (DNN) and obtains physical property values ​​880 of the new chemical structure 850 as output values ​​of the deep neural network (DNN).

[0109] In one embodiment, the chemical structure generation model of the neural network device is trained using the final inverse blacklist vector 840. Specifically, the neural network device decodes the final inverse blacklist vector 840 to generate a new chemical structure 850. The neural network device analyzes the new chemical structure 850 to determine whether the new chemical structure 850 contains a blacklist that is highly likely to exist in the structure input 810. That is, the neural network device calculates a prediction result of the blacklist by analyzing the new chemical structure 850.

[0110] The chemical structure generation model of the neural network device is trained by receiving as feedback a new chemical structure 850 generated from the final inverse blacklist vector 840 and the prediction result of the blacklist corresponding to the new chemical structure 850.

[0111] On the other hand, after the learning of the chemical structure generation model is completed, the neural network device generates a new chemical structure 850 by selecting only a part of the multiple blacklists used in the learning process (hereinafter, the selected blacklists).

[0112] Specifically, the neural network device generates a negative attention vector 820 based on the structure input 810, which represents a respective probability of the presence of each of the selected blacklists in the structure input 810. For example, the neural network device can effectively remove the non-selected blacklists that are not selected by changing the values ​​of elements of the negative attention vector 820 that correspond to the non-selected blacklists to 0.

[0113] The neural network device uses the negative attention vector 820 to calculate the inverse negative attention vector 860. In one embodiment, the neural network device generates a final inverse blacklist vector 840 that represents the portion that does not correspond to the selected blacklist through an element-wise operation of the inverse negative attention vector 860 and the structural representation 830. The neural network device decodes the final inverse blacklist vector 840 to generate a new chemical structure 850.

[0114] As a result, even if the blacklist used to train the chemical structure generation model is subsequently changed (e.g., only a portion of the blacklist is used), the neural network device does not need to re-train the chemical structure generation model each time the blacklist is changed, and can generate a new chemical structure 850 using the already-trained chemical structure generation model.

[0115] FIG. 9 is a diagram illustrating a process of generating a new chemical structure using a final whitelist vector according to one embodiment.

[0116] The neural network device can use not only the blacklist but also the whitelist in the learning process and inference process of the chemical structure generation model. Hereinafter, the description overlapping with FIG. 7 and FIG. 8 will be omitted for convenience.

[0117] Referring to FIG. 9, the neural network device receives a structural input 910 of a chemical structure.

[0118] The neural network device receives the whitelist. The whitelist is the opposite concept of the blacklist and refers to partial structures that are advantageous in synthesis and performance. Specifically, a chemical structure newly generated through the neural network device has a physical property value for a specific physical property (e.g., transmission wavelength, emission wavelength, etc.), and the neural network device aims to generate a new chemical structure having a physical property value that matches a target physical property value (e.g., "emission wavelength: λ=350 nm"). In this case, if the new chemical structure includes a partial structure corresponding to the whitelist, the physical property value of the new chemical structure is more likely to match the target physical property value.

[0119] Based on the structure input 910, the neural network device generates a positive attention vector 920 that represents a respective probability of the presence of each of a plurality of whitelists in the structure input 910.

[0120] The size of the positive attention vector 920 is determined by the number of whitelists. Each element of the positive attention vector 920 has a value between 0 and 1, and each element of the positive attention vector 920 indicates the probability that the corresponding whitelist is present in the structure input 810.

[0121] The neural network device generates a final whitelist vector 940 including a plurality of whitelists through an operation between the positive attention vector 920 and the structural representation 930. In one embodiment, the neural network device generates the final whitelist vector 940 through an element-wise operation between the positive attention vector 920 and the structural representation 930.

[0122] The neural network device decodes the final whitelist vector 940 to generate new chemical structures 950.

[0123] Meanwhile, the chemical structure generation model of the neural network device is trained using the final whitelist vector 940. Specifically, the neural network device generates a new chemical structure 950 by decoding the final whitelist vector 940. The neural network device analyzes the new chemical structure 950 to determine whether the new chemical structure 950 includes a whitelist that is highly likely to be present in the structure input 910. That is, the neural network device calculates a whitelist prediction result 960 by analyzing the new chemical structure 950.

[0124] The chemical structure generation model of the neural network device is trained by receiving as feedback a new chemical structure 950 generated from the final whitelist vector 940 and a whitelist prediction result 960 corresponding to the new chemical structure 950.

[0125] In one embodiment, during the learning process of the chemical structure generation model, the neural network device receives only chemical structures that are confirmed to include at least a portion of the multiple whitelists in the structure input 910. For example, if there are a total of 13 whitelists, the neural network device receives only chemical structures that have at least 10 whitelists out of the 13 whitelists in the structure input 910.

[0126] In addition, during the learning process of the chemical structure generation model, the neural network device receives as structure input 910 only chemical structures that have been confirmed to not include at least a portion of the multiple blacklists and to include at least a portion of the multiple whitelists.

[0127] As described above, the learning performance of the chemical structure generation model of the neural network device is improved by receiving and learning only specific chemical structures in the structure input 810 that include at least a portion of the whitelist or that do not include at least a portion of the blacklist and that include at least a portion of multiple whitelists.

[0128] Fig. 10 is a flow chart for explaining a method for generating a new chemical structure in a neural network device according to an embodiment. Since the method for generating a chemical structure in the neural network device shown in Fig. 10 is related to the embodiment explained in the above drawings, even if the contents are omitted below, the contents explained in the above drawings are also applied to the method in Fig. 10.

[0129] Referring to FIG. 10, in step 1010, the neural network device receives a structural input of a chemical structure.

[0130] In one embodiment, the neural network device receives as structure input a representation of a particular chemical structure, including a fingerprint of the molecular structure (e.g., Morgan Fingerprint (ECFP)) or a quantitative structure-property relationship (QSPR) representation constructed from immediately calculated values ​​such as the molecular weight or the number of substructures (e.g., rings) contained within the molecular structure.

[0131] In step 1020, the neural network device generates a negative attention vector based on the structure input, the negative attention vector representing a respective probability of the presence of each of a plurality of blacklists in the structure input.

[0132] The blacklist refers to substructures that are detrimental to synthesis and performance.

[0133] The size of the negative attention vector is determined by the number of blacklists, and each element of the negative attention vector has a value from 0 to 1, and each element of the negative attention vector indicates the probability that the corresponding blacklist exists in the structure input.

[0134] In step 1030, the neural network device encodes the structural input to generate a structural representation.

[0135] In one embodiment, the neural network device utilizes a CVAE encoder to encode the structural input into a structural representation.

[0136] In step 1040, the neural network device generates a final inverse blacklist vector that does not include multiple blacklists based on the negative attention vector and the structural representation.

[0137] The neural network device calculates an inverse negative attention vector using the negative attention vector. The neural network device generates a final inverse blacklist vector representing a part not corresponding to the plurality of blacklists through an operation between the inverse negative attention vector and the structural representation. For example, the neural network device generates the final inverse blacklist vector through an element-by-element multiplication operation between the inverse negative attention vector and the structural representation.

[0138] In step 1050, the neural network device decodes the final inverse blacklist vector to generate new chemical structures.

[0139] The neural network device uses the CVAE decoder to decode the final blacklist vector into a new chemical structure, which is represented by a representation.

[0140] It is highly likely that the new chemical structure generated based on the final reverse blacklist vector will not contain any part that corresponds to the blacklist for the structure input. The blacklist means a partial structure that is disadvantageous to synthesis and performance, but if the new chemical structure does not contain the blacklist, the physical property value of the new chemical structure is highly likely to match the target physical property value.

[0141] The neural network device inputs a new chemical structure to a recurrent neural network (RNN) and obtains a structural feature value of the new chemical structure as an output value of the recurrent neural network (RNN).The neural network device also inputs the new chemical structure to a deep neural network (DNN) and obtains a physical property value of the new chemical structure as an output value of the deep neural network (DNN).

[0142] Meanwhile, in the learning process of the chemical structure generation model, the neural network device generates a final blacklist vector including a plurality of blacklists through an operation between the negative attention vector and the structural representation. For example, the neural network device generates the final blacklist vector through an element-by-element multiplication operation between the negative attention vector and the structural representation.

[0143] In addition, the neural network device feeds back the prediction results of the new chemical structure and the blacklist corresponding to the new chemical structure to the chemical structure generation model, thereby training the chemical structure generation model.

[0144] The present embodiment is embodied in the form of a recording medium including computer executable instructions such as a program module executed by a computer. The computer readable recording medium is any available medium that can be accessed by a computer, and includes both volatile and non-volatile media, and both separate and non-separate media. The computer readable recording medium also includes both computer recording media and communication media. The computer recording medium includes both volatile and non-volatile, separate and non-separate media embodied by any method or technology for storing information such as computer readable instructions, data structures, program modules, or other data. The communication medium typically includes computer readable instructions, data structures, other data in a modulated data signal such as a program module, or other transmission mechanism, and includes any information delivery medium.

[0145] Also, in this specification, a "unit" refers to a hardware configuration such as a processor or a circuit, and / or a software configuration executed by a hardware configuration such as a processor.

[0146] The above description of the present specification is for illustrative purposes only, and a person skilled in the art to which the present specification pertains can easily understand that the present invention can be easily modified into other specific forms without changing the technical idea or essential features of the present invention. Therefore, it should be understood that the above-described embodiments are illustrative in all respects and are not limiting. For example, each component described as a single type may be implemented in a distributed manner, and similarly, each component described as distributed may be implemented in a combined form.

[0147] Although the embodiments of the present invention have been described in detail above with reference to the drawings, the present invention is not limited to the above-described embodiments and can be modified in various ways without departing from the technical spirit of the present invention. [Explanation of symbols]

[0148] 20,500 Deep Neural Networks (DNN) 30, 31, 32, 510 Recurrent Neural Networks (RNN) 40, 520 CVAE (conditional variational autoencoder) 100 Neural network device 110 Processor 120 Memory 500 Deep Neural Networks (DNN) 510 Recurrent Neural Networks (RNN) 610, 810, 910 Structure input 620, 820 Negative Attention Vectors 630, 830, 930 Structural representation 640 Final Blacklist Vectors 650, 850, 950 New chemical structure 660 Blacklist Prediction Results 670, 870, 970 Structural characteristic value 680, 880, 980 Physical properties 840 Final Reverse Blacklist Vectors 860 Reverse Negative Attention Vectors 920 Positive Attention Vectors 940 Final Whitelist Vectors 960 Whitelist Prediction Results

Claims

1. A method for generating new chemical structures using a neural network device, comprising the steps of: receiving a structural input of a chemical structure; generating a negative attention vector based on the structure input, the negative attention vector representing a respective probability of the presence of each of a plurality of blacklists in the structure input; encoding the structural input to generate a structural representation; generating a final reverse blacklist vector that does not include the plurality of blacklists based on the negative attention vector and the structural representation; and decoding the final reverse blacklist vector to generate a new chemical structure; The method is characterized in that the blacklist includes substructures that are detrimental to synthesis and performance as elements constituting the chemical structure.

2. The step of generating the final reverse blacklist vector comprises: calculating an inverse negative attention vector using the negative attention vector; and generating a final reverse blacklist vector that does not include the plurality of blacklists through an operation between the reverse negative attention vector and the structural representation.

3. The step of generating the final reverse blacklist vector includes a step of generating a final blacklist vector including the plurality of blacklists through an operation between the negative attention vector and the structural representation in a learning process of a chemical structure generation model; The method according to claim 1, wherein the step of generating a new chemical structure further includes a step of training the chemical structure generation model by feeding back prediction results of the new chemical structure and a blacklist corresponding to the new chemical structure.

4. The method of claim 3 , wherein the step of training the chemical structure generation model further comprises the step of calculating a prediction result of the blacklist using a non-negative parameter.

5. The method according to claim 1, characterized in that the receiving step includes a step of receiving as the structure input a chemical structure that has been confirmed not to include at least a portion of the plurality of blacklists during a learning process of a chemical structure generation model.

6. The method further includes selecting a portion of the plurality of blacklists; generating the negative attention vector includes generating a negative attention vector based on the structure input, the negative attention vector representing a respective probability of each of the selected blacklists being present in the structure input; 2. The method of claim 1 , wherein generating the final reverse blacklist vector comprises generating a final reverse blacklist vector that does not include the selected blacklist based on the negative attention vector and the structural representation.

7. 3. The method of claim 2, wherein generating the final reverse blacklist vector comprises generating the final reverse blacklist vector via an element-wise multiplication operation of the reverse negative attention vector and the structural representation.

8. The method includes generating a positive attention vector based on the structure input, the positive attention vector representing a respective probability of the presence of each of a plurality of whitelists in the structure input; generating a final whitelist vector based on the positive attention vector and the structural representation, the final whitelist vector including the plurality of whitelists; and decoding the final whitelist vector to generate a new chemical structure. The method of claim 1 , wherein the whitelist includes substructures that are advantageous for synthesis and performance as elements that make up the chemical structure.

9. The method according to claim 8, further comprising a step of receiving as the structure input a chemical structure that has been confirmed to include at least a portion of the plurality of whitelists during a training process of a chemical structure generation model.

10. The method according to claim 8, further comprising a step of receiving as the structure input a chemical structure that has been confirmed to not include at least a portion of the plurality of blacklists and to include at least a portion of the plurality of whitelists during a learning process of a chemical structure generation model.

11. A neural network device for generating new chemical structures, comprising: a memory having at least one program stored therein; a processor for driving the neural network; The processor executes the at least one program to receiving a structural input of a chemical structure; generating a negative attention vector based on the structure input, the negative attention vector representing a respective probability of the presence of each of a plurality of blacklists in the structure input; encoding the structural input to generate a structural representation; Generate a final reverse blacklist vector representing a portion that does not correspond to the plurality of blacklists based on the negative attention vector and the structural representation; generating a new chemical structure based on the final reverse blacklist vector; A neural network device, wherein the blacklist includes partial structures that are disadvantageous in synthesis and performance as elements constituting the chemical structure.

12. The processor, Calculating an inverse negative attention vector using the negative attention vector; The neural network device of claim 11, further configured to generate a final inverse blacklist vector that does not include the plurality of blacklists through an operation between the inverse negative attention vector and the structural representation.

13. The processor, In the learning process of the chemical structure generation model, a final blacklist vector including the plurality of blacklists is generated through the operation of the negative attention vector and the structural representation; The neural network device according to claim 12, further configured to train the chemical structure generation model by feeding back prediction results of the new chemical structure and a blacklist corresponding to the new chemical structure.

14. 14. The neural network device of claim 13, wherein the processor is further configured to calculate a predicted outcome of the blacklist utilizing non-negative parameters.

15. The neural network device of claim 11, wherein the processor is further configured to receive as the structure input a chemical structure that has been confirmed not to include at least a portion of the plurality of blacklists during a learning process of a chemical structure generation model.

16. The processor, Selecting a portion of the plurality of blacklists; generating a negative attention vector based on the structure input, the negative attention vector representing a respective probability of each of the selected blacklists being present in the structure input; The neural network apparatus of claim 11, further configured to generate a final inverse blacklist vector that does not include the selected blacklist based on the negative attention vector and the structural representation.

17. The processor, generating a positive attention vector based on the structure input, the positive attention vector representing a respective probability of the presence of each of a plurality of whitelists in the structure input; generating a final whitelist vector representing a portion of the plurality of whitelists based on the positive attention vector and the structural representation; further configured to decode the final whitelist vector to generate a new chemical structure; The neural network device according to claim 11 , wherein the whitelist includes substructures that are advantageous in synthesis and performance as elements that constitute the chemical structure.

18. The neural network device of claim 17, wherein the processor is further configured to receive as the structure input a chemical structure that has been confirmed to include at least a portion of the plurality of whitelists during a learning process of a chemical structure generation model.

19. The neural network device of claim 17, further configured to receive as the structure input a chemical structure that has been confirmed to not include at least a portion of the plurality of blacklists and to include at least a portion of the plurality of whitelists during a learning process of a chemical structure generation model.

20. A computer-readable recording medium having recorded thereon a program for causing a computer to execute the method according to any one of claims 1 to 10.

Images