[3N4-GS-10] AI application: model

[3N4-GS-10-05] Molecular Descriptors Based on Global Structure Information of Substructures

〇Lisa Hamada¹, Seiji Takeda¹, Akihiro Kishimoto¹, Hsianghan Hsu¹, Daiju Nakano¹ (1. IBM)

Keywords:Materials Informatics

In Cheminformatics, molecular descriptors are widely used in quantitative structure-property relationships (QSPR) to describe the structural features of molecules and evaluate their contributions to chemical properties. Although various molecular descriptors have been developed so far, most of them consider only the local information of molecules such as counting specific atoms or substructures. Meanwhile, the chemical properties of molecules are strongly influenced by intramolecular interactions, which depends on the positional relationships between substructures. We present new molecular descriptors based on the topological distance between substructures, thus implicitly allowing to account for intramolecular interactions. Our empirical results show that a prediction model of physical property yields better accuracy with the feature vector based on our new descriptor method than other well-known methods.