US-12626787-B2 - Method and apparatus for generating new chemical structure using neural network

US12626787B2US 12626787 B2US12626787 B2US 12626787B2US-12626787-B2

Abstract

A neural network apparatus for generating a new chemical structure may receive a structure input of a chemical structure; generate, based on the structure input, 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, based on the negative attention vector and the structure expression; and generate the new chemical structure by decoding the final reverse blacklist vector.

Inventors

Youngchun KWON
Jiho Yoo
Younsuk Choi
Youngmin NAM
Minsik Park
Jinwoo Park
Dongseon LEE

Assignees

SAMSUNG ELECTRONICS CO., LTD.

Dates

Publication Date: 20260512
Application Date: 20201208
Priority Date: 20191209

Claims (15)

1 . A processor-implemented method of a neural network apparatus that includes a processor and a memory storing relationship between descriptor information, property information, and structural formula information, the method comprising: obtaining a trained chemical structure generation model by operating a conditional variational autoencoder (CVAE), a deep neural network (DNN) and a recurrent neural network (RNN) together, the CVAE comprising an encoder and a decoder; receiving, by the encoder of the CV AE, a high-dimensional descriptor of a structure input of a chemical structure as structure information; performing, by the encoder of the CV AE, an encoding operation to convert the high-dimensional descriptor into a lower-dimensional latent variable; performing, by the decoder of the CV AE, a decoding operation to decode the lower-dimensional latent variable to generate descriptor information, which is high-dimensional data corresponding to a new chemical structure expression; performing a first training operation to train the DNN based on the descriptor information and property information to obtain a factor defining a relationship between the descriptor information and the property information, and performing a second training operation to train the RNN based on the structure information and at least one of the descriptor information or the factor obtained in the first training operation; generating, based on the structure input, a negative attention vector comprising a plurality of first elements, each of the plurality of first elements corresponding to one of a plurality of blacklists, and each of the plurality of first elements indicating a probability that the corresponding one of the plurality of blacklists is present in the structure input; generating a reverse negative attention vector based on a difference between a fundamental vector and the negative attention vector, the fundamental vector representing a vector having a same size as the negative attention vector and having elements, all of which have a value of 1; generating a final reverse blacklist vector based on an element-wise multiplication between the reverse negative attention vector and the structure expression, the final reverse blacklist vector not including the plurality of blacklists; generating, by decoding the final reverse blacklist vector using the decoder of the trained chemical structure generation model, a new chemical structure expression that does not include the plurality of blacklists; generating a final blacklist vector comprising the plurality of blacklists based on the negative attention vector and the structure expression during the obtaining of the trained chemical structure generation model; iteratively performing the first and the second training operations to obtain an updated trained chemical structure generation model based on the new chemical structure expression and a blacklist prediction result corresponding to the new chemical structure expression; and outputting the new chemical structure expression that does not include the plurality of blacklists, wherein the DNN comprises a plurality of layers including an input layer, one or more hidden layers and an output layer, wherein each of the plurality of layers comprises a plurality of channels, wherein each of the plurality of channels comprises a processing element, wherein the descriptor information is provided in the input layer, the property information is provided in the output layer, the factor may be provided in the one or more hidden layers, wherein the structure information comprises a character string, wherein, in each step of the RNN, a subsequent part of the character string, following a current part of the character string input at a time t, is used as an input at a time t+1, wherein the RNN is trained using the factor obtained from the DNN so that the new chemical structure expression reflects the relationship between the descriptor information and the property information defined by the DNN.
2 . The method of claim 1 , further comprising: calculating the blacklist prediction result using a non-negative parameter.
3 . The method of claim 1 , wherein the structure input includes chemical structures confirmed not to include at least one of the plurality of blacklists during a learning process of the trained chemical structure generation model.
4 . The method of claim 1 , further comprising: selecting one or more of the plurality of blacklists, wherein the generating of the negative attention vector comprises generating, based on the structure input, the negative attention vector in which the respective probability of presence of each of the selected blacklists in the structure input is indicated, and wherein the generating of the final reverse blacklist vector comprises generating the final reverse blacklist vector that does not include the selected blacklists based on the negative attention vector and the structure expression.
5 . The method of claim 1 , further comprising: generating, based on the structure input, a positive attention vector comprising a plurality of second elements, each of the plurality of second elements corresponding to one of a plurality of whitelists, and each of the plurality of second elements indicating a probability that the corresponding one of the plurality of whitelists is present in the structure input; generating a final whitelist vector including the plurality of whitelists based on the positive attention vector and the structure expression; and generating the new chemical structure by decoding the final whitelist vector.
6 . The method of claim 5 , wherein the structure input includes chemical structures confirmed to include at least one of the plurality of whitelists during a learning process of the trained chemical structure generation model.
7 . The method of claim 5 , wherein the structure input includes chemical structures confirmed not to include at least one of the plurality of blacklists, and confirmed to include at least one of the plurality of whitelists during a learning process of the trained chemical structure generation model.
8 . A neural network apparatus for generating a new chemical structure expression, the neural network apparatus comprising: a memory configured to store relationship between descriptor information, property information, and structural formula information; and a processor configured to: obtain a trained chemical structure generation model by operating a conditional variational autoencoder (CVAE), a deep neural network (DNN) and a recurrent neural network (RNN) together, the CVAE comprising an encoder and a decoder; receive, by the encoder of the CV AE, a high-dimensional descriptor of a structure input of a chemical structure as structure information; perform, by the encoder of the CV AE, an encoding operation to convert the high-dimensional descriptor into a lower-dimensional latent variable; perform, by the decoder of the CV AE, a decoding operation to decode the lower-dimensional latent variable to generate descriptor information, which is high-dimensional data corresponding to a new chemical structure expression; perform a first training operation to train the DNN based on the descriptor information and property information to obtain a factor defining a relationship between the descriptor information and the property information, and perform a second training operation to train the RNN based on the structure information and at least one of the descriptor information or the factor obtained in the first training operation; generate, based on the structure input, a negative attention vector comprising a plurality of first elements, each of the plurality of first elements corresponding to one of a plurality of blacklists, and each of the plurality of first elements indicating a probability that the corresponding one of the plurality of blacklists is present in the structure input; generating a reverse negative attention vector based on a difference between a fundamental vector and the negative attention vector, the fundamental vector representing a vector having a same size as the negative attention vector and having elements, all of which have a value of 1; generate a final reverse blacklist vector based on an element-wise multiplication between the reverse negative attention vector and the structure expression, the final reverse blacklist vector not including the plurality of blacklists; generate, by decoding the final reverse blacklist vector using the decoder of the trained chemical structure generation model, a new chemical structure expression that does not include the plurality of blacklists; generate a final blacklist vector comprising the plurality of blacklists based on the negative attention vector and the structure expression during the obtaining of the trained chemical structure generation model; iteratively perform the first and the second training operations to obtain an updated trained chemical structure generation model based on the new chemical structure expression and a blacklist prediction result corresponding to the new chemical structure expression; and output the new chemical structure expression that does not include the plurality of blacklists, wherein the DNN comprises a plurality of layers including an input layer, one or more hidden layers and an output layer, wherein each of the plurality of layers comprises a plurality of channels, wherein each of the plurality of channels comprises a processing element, wherein the descriptor information is provided in the input layer, the property information is provided in the output layer, the factor may be provided in the one or more hidden layers, wherein the structure information comprises a character string, wherein, in each step of the RNN, a subsequent part of the character string, following a current part of the character string input at a time t, is used as an input at a time t+1, wherein the RNN is trained using the factor obtained from the DNN so that the new chemical structure expression reflects the relationship between the descriptor information and the property information defined by the DNN.
9 . The neural network apparatus of claim 8 , wherein the processor is further configured to: calculate the blacklist prediction result using a non-negative parameter.
10 . The neural network apparatus of claim 8 , wherein the structure input includes chemical structures confirmed not to include at least one of the plurality of blacklists during a learning process of the trained chemical structure generation model.
11 . The neural network apparatus of claim 8 , wherein the processor is further configured to: select one or more of the plurality of blacklists; generate, based on the structure input, the negative attention vector that indicates a respective probability of presence of each of the selected blacklists in the structure input; and generate the final reverse blacklist vector that does not include the selected blacklists based on the negative attention vector and the structure expression.
12 . The neural network apparatus of claim 8 , wherein the processor is further configured to: generate, based on the structure input, a positive attention vector comprising a plurality of second elements, each of the plurality of second elements corresponding to one of a plurality of whitelists, and each of the plurality of second elements indicating a probability that the corresponding one of the plurality of whitelists is present in the structure input; generate a final whitelist vector that indicates portions corresponding to the plurality of whitelists, based on the positive attention vector and the structure expression; and generate the new chemical structure by decoding the final whitelist vector.
13 . The neural network apparatus of claim 12 , wherein the structure input includes chemical structures confirmed to include at least one of the plurality of whitelists during a learning process of the trained chemical structure generation model.
14 . The neural network apparatus of claim 12 , wherein the structure input includes chemical structures confirmed not to include at least one of the plurality of blacklists and confirmed to include at least one of the plurality of whitelists during a learning process of the trained chemical structure generation model.
15 . A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the method of claim 1 .

Description

CROSS-REFERENCE TO RELATED APPLICATION This application is based on and claims priority under 35 U.S.C. § 119 to Korean Patent Application No. 10-2019-0162918, filed on Dec. 9, 2019, in the Korean Intellectual Property Office, the disclosure of which is incorporated by reference herein in its entirety. BACKGROUND 1. Field The present disclosure relates to an apparatus and method for generating a new chemical structure. 2. Description of Related Art Neural networks refer to a computational architecture in which a biological brain is modeled. With the development of neural network technology, neural networks have been used in various types of electronic systems to analyze input data and extract effective information. Recently, research has been actively carried out to evaluate properties of chemical structures by using neural network technology so as to select chemical structures to be used for developing materials. According to a method of designing a material using neural network technology according to the related art, it is checked ex post facto whether or not a new chemical structure includes predetermined blacklists after obtaining the new chemical structure, and thus such a new chemical structure including the blacklists is filtered. However, most new chemical structures generated using a neural network model include the blacklists, and thus, it is difficult to obtain new chemical structures having satisfactory properties. SUMMARY Provided are apparatuses and methods for generating a chemical structure using a neural network. Provided is a computer-readable recording medium in which a program for executing the method on a computer is recorded. Technical problems to be solved are not limited to the above technical problems, and thus other technical problems may exist. 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 of the disclosure. According to an aspect of an embodiment, a method of generating a new chemical structure may be performed by a neural network apparatus and may include receiving a structure input; generating, based on the structure input, a negative attention vector that indicates a respective probability of presence of each of a plurality of blacklists in the structure input; generating a structure expression by encoding the structure input; generating a final reverse blacklist vector that does not include the plurality of blacklists, based on the negative attention vector and the structure expression; and generating the new chemical structure by decoding the final reverse blacklist vector. The method may include calculating a reverse negative attention vector using the negative attention vector, and generating of the final reverse blacklist vector may include generating the final reverse blacklist vector based on the reverse negative attention vector and the structure expression. The method may include generating a final blacklist vector that includes the plurality of blacklists based on the negative attention vector and the structure expression during a learning process of a chemical structure generation model; and training the chemical structure generation model based on the new chemical structure and a blacklist prediction result corresponding to the new chemical structure. The method may include calculating the blacklist prediction result using a non-negative parameter. The structure input may include chemical structures confirmed not to include at least a part of the plurality of blacklists during a learning process of a chemical structure generation model. The method may include selecting a part of the plurality of blacklists. The generating of the negative attention vector may include generating, based on the structure input, the negative attention vector in which the respective probability of presence of each of the selected blacklists in the structure input is indicated, and the generating of the final reverse blacklist vector may include generating the final reverse blacklist vector that does not include the selected blacklists based on the negative attention vector and the structure expression. The generating of the final reverse blacklist vector may include generating the final reverse blacklist vector based on an element-wise multiplication between the reverse negative attention vector and the structure expression. The method may include generating, based on the structure input, a positive attention vector that indicates a respective probability of presence of each of a plurality of whitelists in the structure input; generating a final whitelist vector including the plurality of whitelists based on the positive attention vector and the structure expression; and generating the new chemical structure by decoding the final whitelist vector. The structure input may include chemical structures confirmed to include at least a part of the plu