KFU. Card publication. Deep conditional variational autoencoder for reaction conditions prediction

KAZAN
FEDERAL UNIVERSITY

- RUS
- 中文
- ES

DEEP CONDITIONAL VARIATIONAL AUTOENCODER FOR REACTION CONDITIONS PREDICTION

Form of presentation

Conference proceedings in international journals and collections

Year of publication

2020

Язык

русский

Authors, employees of KFU

Afonina Valentina Aleksandrovna, author

Varnek Aleksandr , author

Lin Arkadiy Igorevich, author

Madzhidov Timur Ismailovich, author

Mazitov Daniyar Amirovich, author

Nugmanov Ramil Irekovich, author

Skibina Yuliya Dmitrievna, author

Authors, students KFU

Akhmetshin Tagir Nailevich, author

Bibliographic description in the original language

Madzhidov T.I. Deep conditional variational autoencoder for reaction conditions prediction / T.I.Madzhidov, D.A. Mazitov, V.A. Afonina, T.N. Akhmetshin, J.D. Skibina, A.I. Lin, R.I. Nugmanov, A. Varnek // 3rd RSC-BMCS/RSC-CICAG Artificial Intelligence in Chemistry. - London, 2020. - P. 55.

Annotation

3rd RSC-BMCS/RSC-CICAG Artificial Intelligence in Chemistry

Keywords

хемоинформатика, искусственный интеллект, предсказание условий, информатика реакций

The name of the journal

3rd RSC-BMCS/RSC-CICAG Artificial Intelligence in Chemistry

Please use this ID to quote from or refer to the card

https://repository.kpfu.ru/eng/?p_id=258451&p_lang=2

Resource files

File name	Size (MB)	Format
thesis.pdf	0,17	pdf	show / download

Full metadata record

Field DC	Value	Language
dc.contributor.author	Afonina Valentina Aleksandrovna	ru_RU
dc.contributor.author	Varnek Aleksandr	ru_RU
dc.contributor.author	Lin Arkadiy Igorevich	ru_RU
dc.contributor.author	Madzhidov Timur Ismailovich	ru_RU
dc.contributor.author	Mazitov Daniyar Amirovich	ru_RU
dc.contributor.author	Nugmanov Ramil Irekovich	ru_RU
dc.contributor.author	Skibina Yuliya Dmitrievna	ru_RU
dc.contributor.author	Akhmetshin Tagir Nailevich	ru_RU
dc.date.accessioned	2020-01-01T00:00:00Z	ru_RU
dc.date.available	2020-01-01T00:00:00Z	ru_RU
dc.date.issued	2020	ru_RU
dc.identifier.citation	Madzhidov T.I. Deep conditional variational autoencoder for reaction conditions prediction / T.I.Madzhidov, D.A. Mazitov, V.A. Afonina, T.N. Akhmetshin, J.D. Skibina, A.I. Lin, R.I. Nugmanov, A. Varnek // 3rd RSC-BMCS/RSC-CICAG Artificial Intelligence in Chemistry. - London, 2020. - P. 55.	ru_RU
dc.identifier.uri	https://repository.kpfu.ru/eng/?p_id=258451&p_lang=2	ru_RU
dc.description.abstract	3rd RSC-BMCS/RSC-CICAG Artificial Intelligence in Chemistry	ru_RU
dc.description.abstract	Synthesis planning domain has a second birthday after seminal paper of Segler et al.1 who proposed AI-powered technology for retrosynthetic route prediction. However, selection of optimal conditions for every reaction in synthetic route is almost as challenging and as important as prediction of the route itself. The main complication in prediction of optimal conditions are absence of negative examples, very large condition space, voids in reaction-condition matrix (all possible conditions for given reaction are unavailable). For some reactions, conditions are well-studied and documented, but for the other selection of conditions could be tricky and requires good knowledge and experience of chemists. Here, we propose an approach based on deep neural networks that predicts combination of catalyst-reagent-temperature-pressure best suited for a particular reaction. Unlike existing approaches2,3, here we do not rank possible conditions based on some technique but directly sample them using conditional variational autoencoder. Training and test sets comprised of hydrogenation reactions from Reaxys database, two datasets are collected: “small” consisting from 38K with limited set of conditions and “big” one including 196K reactions with vast variety of conditions. Three different latent distributions were compared, namely Gaussian (g-CVAE), Riemannian Normalizing Flow (mf-CVAE) and Hyperspherical Uniform (h-CVAE) distributions. Proposed approaches were shown superior performance to null model, which predicts conditions based on their popularity in training data, as well as over nearest-neighbor approach (KNN). The latter ranks possible conditions based on the similarity of corresponding reactions to a given one. For hydrogenation reactions, previously proposed approach of Gao et al.2 have shown mediocre performance	ru_RU
dc.language.iso	ru	ru_RU
dc.subject	хемоинформатика	ru_RU
dc.subject	искусственный интеллект	ru_RU
dc.subject	предсказание условий	ru_RU
dc.subject	информатика реакций	ru_RU
dc.title	Deep conditional variational autoencoder for reaction conditions prediction	ru_RU
dc.type	Conference proceedings in international journals and collections	ru_RU