XGBoost, mordred and RDKit for the prediction of glass transition temperature of polymers

Glass transition temperature (Tg) is the temperature at which a polymer changes from crystalline state to rubbery state. This change in the property below and above Tg is very important in food science and pharmaceutical industries. In recent decades, there has been a growth in using machine learni...

全面介紹

Saved in:
書目詳細資料
主要作者: Goh, Kai Leong
其他作者: Lu Yunpeng
格式: Student Research Paper
語言:English
出版: Nanyang Technological University 2022
主題:
在線閱讀:https://hdl.handle.net/10356/155298
標簽: 添加標簽
沒有標簽, 成為第一個標記此記錄!
機構: Nanyang Technological University
語言: English
實物特徵
總結:Glass transition temperature (Tg) is the temperature at which a polymer changes from crystalline state to rubbery state. This change in the property below and above Tg is very important in food science and pharmaceutical industries. In recent decades, there has been a growth in using machine learning (ML) to develop quantitative structure–property relationship (QSPR) models. QSPR uses molecular descriptors and molecular fingerprints as features to predict the properties of chemical compounds. As a result, numerous works have been dedicated to creating a good QSPR model to predict Tg. However, to the best of our knowledge, there was no previous research work that involved the use of the Mordred molecular descriptors library or the Extreme Gradient Boosting (XGBoost) regression algorithm to predict Tg. Therefore, this project employed Mordred and XGBoost, together with the RDKit cheminformatics library to predict Tg of 640 polymers. A total of 12 sets of features were generated by RDKit and Mordred as inputs for XGBoost to predict Tg. The scoring metrics from the Scikit-learn and Numpy libraries showed that the 2D molecular descriptors of Mordred (Mordred-2D) and the Extended-Connectivity Fingerprint with a diameter of 4 bonds (ECFP4) had the best performances. The results further improved when Mordred-2D and ECFP4 were combined to form a new set of features. Future work aims to increase the number of polymer data points and explore better methods to represent the polymer repeating units for the calculation of descriptors and fingerprints.