TDiMS : A Topological Distance based Intra-Molecular Substructure Descriptor for Improved Machine Learning Predictions

Abstract

Various molecular descriptors have been developed due to their diverse roles and importance in material informatics. However, they still have challenges in accurately capturing the global relationship of intra-molecular substructures, which significantly influence on the physical property. In this paper, we introduced a novel molecular descriptor which can extract topological distance between each pair of substructures within a molecule. Our evaluations reveal that the proposed descriptor outperformed existing baselines in downstream tasks, including neural-network-based models. Moreover, this descriptor enables to acquire important chemical insight into what substructure pairs need to be considered with topological distance, which is crucial for advanced tasks such as molecular generation.