Property Prediction without Descriptors

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, finding the best set of descriptors for each prediction is very difficult, and the descriptors may not contain sufficient information about the object property. In this study, molecular structures were represented by atom position, atom kind and graph structure. And regression model was constructed using Graph Convolutional Neural Network (GCNN). As a result of the case study, the proposed method outperformed the existing method which use descriptors in a case. But in another case, the proposed method performed worse than existing method. It can be thought that one of the reasons was the insufficiency of representability of model. Thus, the consideration of input form or model structure may improve the prediction ability of the proposed method.