数量结构-活动关系
可解释性
计算机科学
人工智能
机器学习
试验装置
化学信息学
编码(集合论)
卷积神经网络
集合(抽象数据类型)
生物信息学
生物
程序设计语言
作者
J. Tong,Peng Gao,Hai-Yin Xu,Yuan Liu
标识
DOI:10.1016/j.molliq.2023.123708
摘要
In supervised model training, QSAR modeling of drug analogous texture is commonly used, mostly using RNN for learning structural information of molecules, but a problem is a large time cost in fitting temporal data by recurrent training units. In this paper, we propose a novel molecular canonicalization method, GAFMol, leveraging the Gramian Angular Field (GAF) algorithm and Extended-Connectivity Fingerprint (ECFP) algorithm. GAFMol transforms molecular sequences into graphs, providing an innovative approach to capturing structural information. We apply the GAFMol-CNN scheme to construct a robust QSAR model for SARS-CoV-2 Mpro inhibitors, and thoroughly investigate the model's interpretability. Our method demonstrates superior performance compared to existing QSAR models, as discussed in detail in the Results and Discussion section. Compared with previous QSAR models (Monte Carlo, Mol2vec, FCNN, LSTM, bi-directional LSTM and CNN methods), the GAFMol-CNN scheme has significant improvement, and achieved relatively good results in the test set. The source code of the molecular canonicalization method GAFMol is freely available on https://github.com/Xuhaiyin/.
科研通智能强力驱动
Strongly Powered by AbleSci AI