Development of modifications of semi-template based models for predicting possible ways of synthesizing chemicals using deep neural networks
Jazyk: | ruština |
---|---|
Rok vydání: | 2022 |
Předmět: |
predictive chemistry
гÑаÑовÑе нейÑоннÑе ÑеÑи ÑеÑÑоÑинÑез data analysis graph neural networks transformers semi-template based глÑбокие нейÑоннÑе ÑеÑи python пÑедикÑÐ¸Ð²Ð½Ð°Ñ Ñ Ð¸Ð¼Ð¸Ñ machine learning deep neural networks анализ даннÑÑ retrosynthesis маÑинное обÑÑение ÑÑанÑÑоÑмеÑÑ |
DOI: | 10.18720/spbpu/3/2023/vr/vr23-516 |
Popis: | Тема вÑпÑÑÐºÐ½Ð¾Ð¸Ì ÐºÐ²Ð°Ð»Ð¸ÑикаÑÐ¸Ð¾Ð½Ð½Ð¾Ð¸Ì ÑабоÑÑ: «РазÑабоÑка модиÑикаÑÐ¸Ð¸Ì semi-template based Ð¼Ð¾Ð´ÐµÐ»ÐµÐ¸Ì Ð¿ÑедÑÐºÐ°Ð·Ð°Ð½Ð¸Ñ Ð²Ð¾Ð·Ð¼Ð¾Ð¶Ð½ÑÑ Ð¿ÑÑÐµÐ¸Ì ÑинÑеза Ñ Ð¸Ð¼Ð¸ÑеÑÐºÐ¸Ñ Ð²ÐµÑеÑÑв Ñ Ð¸ÑполÑзованием глÑÐ±Ð¾ÐºÐ¸Ñ Ð½ÐµÐ¸ÌÑоннÑÑ ÑеÑеиÌ».ÐÐ°Ð½Ð½Ð°Ñ ÑабоÑа поÑвÑÑена иÑÑÐ»ÐµÐ´Ð¾Ð²Ð°Ð½Ð¸Ñ Ð¼ÐµÑодов ÑеÑаÑÑÐ¸Ñ Ð·Ð°Ð´Ð°ÑÑ Ð¿ÑедÑÐºÐ°Ð·Ð°Ð½Ð¸Ñ Ð²Ð¾Ð·Ð¼Ð¾Ð¶Ð½ÑÑ Ð¿ÑÑÐµÐ¸Ì ÑеÑÑоÑинÑеза Ñ Ð¸Ð¼Ð¸ÑеÑÐºÐ¸Ñ ÑÐ¾ÐµÐ´Ð¸Ð½ÐµÐ½Ð¸Ð¸Ì Ð¸ ÑазÑабоÑке модиÑикаÑÐ¸Ð¸Ì semi-template based меÑодов Ð´Ð»Ñ Ð¿Ð¾Ð²ÑÑÐµÐ½Ð¸Ñ ÑазнообÑÐ°Ð·Ð¸Ñ ÑезÑлÑÑаÑов генеÑиÑÑемÑÑ Ð¼Ð¾Ð´ÐµÐ»ÑÑ.ÐадаÑи, коÑоÑÑе ÑеÑалиÑÑ Ð² Ñ Ð¾Ð´Ðµ ÑабоÑÑ:1. анализ ÑÑÑеÑÑвÑÑÑÐ¸Ñ Ð¼ÐµÑодов и Ð¿Ð¾Ð´Ñ Ð¾Ð´Ð¾Ð² пÑедикÑÐ¸Ð²Ð½Ð¾Ð¸Ì Ñ Ð¸Ð¼Ð¸Ð¸ Ð´Ð»Ñ ÑеÑÐµÐ½Ð¸Ñ Ð·Ð°Ð´Ð°Ñи пÑедÑÐºÐ°Ð·Ð°Ð½Ð¸Ñ Ð¿ÑÑÐµÐ¸Ì ÑеÑÑоÑинÑеза;2. ÑазÑабоÑка модиÑикаÑииÌ, коÑоÑÑе позволÑÑÑ ÑвелиÑиÑÑ ÑазнообÑазие пÑедÑÐºÐ°Ð·Ð°Ð½Ð¸Ñ Ð´Ð»Ñ semi-template based меÑодов3. иÑÑледование ÑÑÑекÑивноÑÑи ÑеализованнÑÑ Ð¼Ð¾Ð´Ð¸ÑикаÑÐ¸Ð¸Ì Ð¼ÐµÑодов.РабоÑа вÑполнена Ñ Ð¸ÑполÑзование оÑкÑÑÑÑÑ Ð½Ð°Ð±Ð¾Ñов даннÑÑ Â Ñ Ð¸Ð¼Ð¸ÑеÑÐºÐ¸Ñ ÑеакÑÐ¸Ð¸Ì USPTO-50K и USPTO-full. Ðа ÑзÑке пÑогÑаммиÑÐ¾Ð²Ð°Ð½Ð¸Ñ Python 3.7 c иÑполÑзованием ÑÑÑиÌмвоÑков Pytorch и Open-NMT бÑли ÑÐµÐ°Ð»Ð¸Ð·Ð¾Ð²Ð°Ð½Ñ Ð¼Ð¾Ð´Ð¸ÑикаÑии Ð´Ð»Ñ semi-template based меÑодов RetroXpert и GraphRetro, коÑоÑÑе ÑлÑÑÑаÑÑ Ð¿Ð¾ÐºÐ°Ð·Ð°Ñели пÑедложеннÑÑ Ð¼ÐµÑÑик Ð´Ð»Ñ Ð¾Ñенки ÑазнообÑÐ°Ð·Ð¸Ñ Ð¿ÑедÑказанииÌ, Diversity и Class Diversity.РиÑоге вÑполнено иÑÑледование ÑеализованнÑÑ Ð¼Ð¾Ð´Ð¸ÑикаÑÐ¸Ð¸Ì Ñ ÑоÑки зÑÐµÐ½Ð¸Ñ Ð¿ÑедложеннÑÑ Ð¼ÐµÑÑик. Ðа оÑновании иÑÑÐ»ÐµÐ´Ð¾Ð²Ð°Ð½Ð¸Ð¸Ì Ð¼Ð¾Ð¶Ð½Ð¾ ÑделаÑÑ Ð²Ñвод,ÑÑо ÑазÑабоÑаннÑе модиÑикаÑии позволÑÑÑ Ð¿Ð¾Ð²ÑÑиÑÑ Ð·Ð½Ð°Ñение меÑÑики Diversity на 2% и Class Diversity 3% Ñ Ð¿Ð¾Ð¼Ð¾ÑÑÑ Ð¿ÐµÑÐ²Ð¾Ð¸Ì Ð¼Ð¾Ð´Ð¸ÑикаÑии и Diversity на 3% и Class Diversity 1.3% Ñ Ð¿Ð¾Ð¼Ð¾ÑÑÑ Ð²ÑоÑÐ¾Ð¸Ì Ð¼Ð¾Ð´Ð¸ÑикаÑии. The subject of the graduate qualification work: "Development of modifications of semi-template based models for predicting possible ways of synthesizing chemicals using deep neural networks."The given work is devoted to the study of methods that solve the problem of predicting possible pathways for the retrosynthesis of chemical compounds and the development of modifications of semi-template based methods to increase the diversity of results generated by the model.The research set the following goals:1. analysis of existing methods and approaches of predictive chemistry for solving the problem of predicting retrosynthesis pathways;2. development of modifications that allow increasing the variety of predictions for semi-template based methods3. research about the performance of implemented modifications of methods.The work was performed using USPTO-50K and USPTO-full open chemical reaction datasets. In the Python 3.7 programming language using the Pytorch and Open-NMT frameworks, modifications were implemented for the semi-template based methods RetroXpert and GraphRetro, which improve the performance of the proposed metrics for assessing the diversity of predictions, Diversity and Class Diversity.As a result, a research of the implemented modifications was carried out in terms of the proposed metrics. Based on the research, it can be concluded that the developed modifications allow increasing the value of the Diversity by 2% and Class Diversity 3% with the first modification and Diversity by 3% and Class Diversity 1.3% with the second modification. |
Databáze: | OpenAIRE |
Externí odkaz: |