Machine learning algorithms for predicting atomic charges and geometry of molecules
File(s)
Author(s)
Xie, Qin
Type
Thesis
Abstract
Drug development is a lengthy and complex process, often involving high costs and long timelines. In recent years, modern machine learning (ML) techniques have shown great promise in accelerating this process, particularly in virtual screening. However, many existing ML models still face limitations in accuracy, scalability, and computational efficiency.
This thesis introduces two original algorithms---MPF (message-passing featuriser) and GLaCE (Graph Latent Connectivity Enhancement)---designed to address these limitations in molecular properties prediction tasks. The MPF algorithm demonstrates strong robustness in atomic charge prediction across various charge schemes. The GLaCE algorithm reliably predicts 2D molecular structures based solely on connectivity information. However, the models perform less well on datasets consisting of molecules that are larger than those in the training set. Both algorithms are designed to be highly flexible, accepting SMILES strings of arbitrary size, and computationally affordable, making them accessible on standard personal computers.
The development and validation of MPF and GLaCE provide a solid methodological foundation for future research in molecular machine learning and offer practical tools for enhancing the efficiency and accessibility of virtual screening in drug discovery.
This thesis introduces two original algorithms---MPF (message-passing featuriser) and GLaCE (Graph Latent Connectivity Enhancement)---designed to address these limitations in molecular properties prediction tasks. The MPF algorithm demonstrates strong robustness in atomic charge prediction across various charge schemes. The GLaCE algorithm reliably predicts 2D molecular structures based solely on connectivity information. However, the models perform less well on datasets consisting of molecules that are larger than those in the training set. Both algorithms are designed to be highly flexible, accepting SMILES strings of arbitrary size, and computationally affordable, making them accessible on standard personal computers.
The development and validation of MPF and GLaCE provide a solid methodological foundation for future research in molecular machine learning and offer practical tools for enhancing the efficiency and accessibility of virtual screening in drug discovery.
Version
Open Access
Date Issued
2024-05-16
Date Awarded
01/09/2025
Advisor
Horsfield, Andrew
Walsh, Aron
Publisher Department
Department of Materials
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)