A Graph Neural Network for Predicting the Magnetic Interaction Between Atoms

Singh, Jaswinder (2021) A Graph Neural Network for Predicting the Magnetic Interaction Between Atoms. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration manual) Download (2MB) | Preview

Abstract

The magnetic interaction between a pair of atoms can be determined by calculating the value of the quantity known as the scalar coupling constant (SCC). The SCC plays a crucial role in the analysis of 3D structure of organic matter and its precise calculation can be used in a variety of tasks like drug discovery, toxicity determination, etc. The quantum mechanical density functional theory (DFT) provides a theoretical framework for predicting the magnetic interactions (or SCC) between the atoms. The quantum mechanical computations use the 3D structural information of the molecule as an input for precise calculation of SCC. However, these computations are extremely time consuming and computationally expensive. To compute SCC efficiently and accurately, a novel graph neural network (GNN) is proposed, combining the message passing elements from the Message Passing Neural Networks (MPNN) model with the multi-head attention layers as in transformer encoder. The proposed model is named as the Message Passing Molecular Transformer (MPMT). Different bond level and atom level features like cosine angles and dihedral angles were created using the RDkit package, to improve the predictions of the model. To demonstrate and validate the superiority of the proposed model, the CHAMPS dataset consisting of structural information and coupling constant values of around 10 million different molecules collected by University of Bristol was used. The results were evaluated on the basis of the log MAE(mean absolute error) values for each coupling type and the final score for the model was computed by averaging over all the coupling types. The MPMT model was able to achieve a final score of -2.873 which corresponds to the mean absolute error of 0.0565 Hz whereas the MPNN was able to achieve an average final score of -2.19 corresponding to the MAE of 0.111 Hz. Our proposed model was able to outperform the state-of-the-art results for the CHAMPS dataset (Jian et al. (2020)) which is MAE value of 0.096 Hz. This study will eventually benefit to various domains; for e.g it will contribute towards efficient and drug development for different diseases, it can be utilised to improve the current NMR techniques for molecular property prediction and can also be used in materials science industry for enhancing cosmetics’ quality. It will also greatly benefit researchers in studying the molecular structure more efficiently, saving lot of research costs.

Item Type:	Thesis (Masters)
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science Q Science > QA Mathematics > Computer software T Technology > T Technology (General) > Information Technology > Computer software Q Science > QC Physics
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Clara Chan
Date Deposited:	14 Dec 2021 15:55
Last Modified:	14 Dec 2021 15:55
URI:	https://norma.ncirl.ie/id/eprint/5225

Actions (login required)

View Item