Property Predicttion from 3D Coordinates of Atoms

Abstract

Quantitative Structure-Property Relationship (QSPR) is a kind of method to predict properties of compounds. In QSPR, a regression model is constructed from training data consisting of the structure and properties of the compound. In　many cases, molecular descriptors are calculated from structure and are used as input. However, the descriptors may not contain sufficient information about the object property, and 3D structure is difficult to consider. In this study, molecular structures were regarded as sequential data of atom information, and regression model was constructed using recurrent neural network (RNN) to deal with variable length data. It was shown that prediction accuracy improves by normalization of coordinate system and consideration of multiple coordinate systems. As a result of the case study, the proposed method outperformed the existing method for predicting octanol/water partition coefficient. This method is expected to be more useful by eliminating the influence from the data format and considering other conformations.