文献类型：期刊文献

英文题名：Transformer-convolutional neural network for surface charge density profile prediction: Enabling high-throughput solvent screening with COSMO-SAC

作者：Chen, Guzhong[1];Song, Zhen[1,2,3];Qi, Zhiwen[1]

机构：[1]East China Univ Sci & Technol, Sch Chem Engn, State Key Lab Chem Engn, 130 Meilong Rd, Shanghai 200237, Peoples R China;[2]Otto von Guericke Univ, Proc Syst Engn, Univ Pl 2, D-39106 Magdeburg, Germany;[3]Max Planck Inst Dynam Complex Tech Syst, Proc Syst Engn, Sandtorstr 1, D-39106 Magdeburg, Germany

年份：2021

卷号：246

外文期刊名：CHEMICAL ENGINEERING SCIENCE

收录：;EI(收录号：20213410798344);WOS:【SCI-EXPANDED(收录号:WOS:000704401400014)】；

基金：The financial support from National Natural Science Foundation of China (21861132019 and 21776074) is greatly acknowledged.

语种：英文

外文关键词：Deep learning; Molecular fingerprints; Convolutional neural networks; COSMO-SAC; sigma-profile prediction; High-throughput solvent screening

摘要：A deep learning (DL) method for quickly predicting surface charge density profiles (sigma-profile) and cavity volumes (V-COSMO) of molecules for the COSMO-SAC model is developed. The molecular fingerprints are derived from the encoder state of a Transformer model pre-trained on the ChEMBL database, which allows transfer learning from large-scale unlabeled data and improve generalization performance by developing better molecular fingerprints for building models with significantly smaller datasets. Employing the pre-trained molecular fingerprints, a convolutional neural network (CNN) model for the sigma-profile and V-COSMO prediction is trained and tested on the VT-2005 database. The obtained Transformer-CNN model presents superior performance to the GC-COSMO approach and enables the pre-diction of sigma-profile and V-COSMO of millions of molecules in only a few minutes. Taking advantages of the model, a high-throughput solvent screening framework based on COSMO-SAC is further proposed and exemplified by searching sustainable solvent for the deterpenation process of citrus essential oils. (C) 2021 Elsevier Ltd. All rights reserved.