Method and apparatus for synthesizing target products by using neural networks

Patent metadata

Field	Value
Publication number	US-12547889-B2
Application number	US-202117372808-A
Country	US
Kind code	B2
Filing date	Jul 12, 2021
Priority date	Jul 29, 2020
Publication date	Feb 10, 2026
Grant date	Feb 10, 2026

Abstract

An apparatus for generating a target product by using a neural network is configured to: predict candidate reactant combinations for generating the target product by using a pre-trained retrosynthesis prediction model; predict a prediction product with respect to each of the candidate reactant combinations by using a pre-trained reaction prediction model; and determine an experimental priority order of the candidate reactant combinations based on a result of comparing the target product with the prediction product.

First claim

What is claimed is:

1 . A method of synthesizing a target product using a neural network, the method comprising:
predicting candidate reactant combinations for generating the target product based on a pre-trained retrosynthesis prediction model using categorical latent variables; predicting a prediction product with respect to each of the candidate reactant combinations based on a pre-trained reaction prediction model using the categorical latent variables shared with the pre-trained retrosynthesis prediction model; performing a comparison operation to compare a string type representation of the target product with a string type representation of the prediction product; determining an experimental priority order of the candidate reactant combinations based on a result of the comparison operation; and experimentally synthesizing the target product based on the experimental priority order of the candidate reactant combinations, wherein the candidate reactant combinations derived through the pre-trained retrosynthesis prediction model are verified through the pre-trained reaction prediction model to prevent a candidate reactant combination inconsistent with a formula grammar to be output for predicting the prediction product.

2 . The method of claim 1 , further comprising,
receiving test reactant combinations and test products corresponding to the test reactant combinations, and learning, using the pre-trained retrosynthesis prediction model and the pre-trained reaction prediction model, the categorical latent variables including a plurality of classes based on the test reactant combinations and the test products.

3 . The method of claim 2 , wherein the learning of the categorical latent variables includes learning a conditional probability distribution of the categorical latent variables corresponding to an input representing each of the test products.

4 . The method of claim 2 , wherein the learning of the categorical latent variables includes learning the categorical latent variables based on a prediction yield rate provided by a pre-trained yield rate prediction model.

5 . The method of claim 2 , wherein the predicting of the candidate reactant combinations and the predicting of the prediction product is based on the learned categorical latent variables.

6 . The method of claim 1 , wherein the predicting of the candidate reactant combinations comprises:
receiving information corresponding to the target product; predicting categorical latent variables including a plurality of classes based on the information corresponding to the target product; and obtaining information corresponding to the candidate reactant combinations based on the information corresponding to the target product and the categorical latent variables.

7 . The method of claim 6 , wherein the obtaining of the information corresponding to the candidate reactant combinations comprises:
calculating a likelihood for each of the plurality of classes of each of the candidate reactant combinations based on an input representing the target product; calculating a likelihood of an retrosynthesis prediction result corresponding to the input representing to the target product and an input representing the categorical latent variables; and selecting a predetermined number of final candidate reactant combinations based on the likelihood for each class and the likelihood of the retrosynthesis prediction result.

8 . The method of claim 6 , wherein the predicting of the categorical latent variables includes providing an expectation reaction method for generating the target product as an input value of the categorical latent variables.

9 . The method of claim 1 , wherein the predicting of the prediction product comprises:
receiving information corresponding to the candidate reactant combinations; receiving categorical latent variables including a plurality of classes; and obtaining information corresponding to the prediction product with respect to each of the candidate reactant combinations, based on the information corresponding to the candidate reactant combinations and the categorical latent variables.

10 . The method of claim 1 , wherein the determining of the experimental priority order of the candidate reactant combinations comprises:
determining whether the prediction product and the target product correspond to each other, based on an input representing each of the candidate reactant combinations; and determining the experimental priority order of the candidate reactant combinations based on whether the prediction product and the target product correspond to each other.

11 . An apparatus including a neural network, the apparatus comprising:
a memory storing at least one program; and a processor configured to execute the at least one program to:
predict candidate reactant combinations for generating a target product based on a pre-trained retrosynthesis prediction model using categorical latent variables;
predict a prediction product with respect to each of the candidate reactant combinations based on a pre-trained reaction prediction model using the categorical latent variables shared with the pre-trained retrosynthesis prediction model;
perform a comparison operation to compare a string type representation of the target product with a string type representation of the prediction product;
determine an experimental priority order of the candidate reactant combinations based on a result of the comparison operation; and
experimentally synthesize the target product based on the experimental priority order of the candidate reactant combinations,
wherein the candidate reactant combinations derived through the pre-trained retrosynthesis prediction model are verified through the pre-trained reaction prediction model to prevent a candidate reactant combination inconsistent with a formula grammar to be output for predicting the prediction product.

12 . The apparatus of claim 11 , wherein the processor is further configured to learn, using the pre-trained retrosynthesis prediction model and the pre-trained reaction prediction model, the categorical latent variables including a plurality of classes based on test reactant combinations and test products corresponding to the test reactant combinations.

13 . The apparatus of claim 12 , wherein the pre-trained retrosynthesis prediction model and the pre-trained reaction prediction model learn a conditional probability distribution of the categorical latent variables with respect to an input representing each of the test products.

14 . The apparatus of claim 12 , wherein the pre-trained retrosynthesis prediction model and the pre-trained reaction prediction model learn the categorical latent variables based on a prediction yield rate provided by a pre-trained yield rate prediction model.

15 . The apparatus of claim 12 , where the processor is further configured to predict the candidate reactant combinations and the predicting of the prediction product is based on the learned categorical latent variables.

16 . The apparatus of claim 11 , wherein the processor is further configured to execute the at least one program to:
receive information corresponding to the target product; predict categorical latent variables including a plurality of classes based on the information corresponding to the target product; and obtain information corresponding to the candidate reactant combinations based on the information corresponding to the target product and the categorical latent variables.

17 . The apparatus of claim 16 , wherein

Assignees

• Samsung Electronics Co Ltd

Inventors

• Park Minsik
• Kwon Youngchun
• Lee Dongseon
• Choi Younsuk
• Kim Eunji

Classifications

• G16C20/70

  Machine learning, data mining or chemometrics · CPC title

• G16C20/10 Primary

  Analysis or design of chemical reactions, syntheses or processes · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title