Solubility prediction using neural network and chemical explanation of deep learning model

Abstract

Solubility is one of the important parameters for designing new compounds. However, it is difficult to get solubility of all candidate compound, so quantitative structure-property relationships (QSPR) is widely used. In this study, we use the deep learning, which is popular in the field of informatics. Deep learning model can represent complex non-linear relationship between the descriptor and the property. To simulate the chemical phenomena, we divided the model in two parts. One is feature extraction part, which converts descriptors into solubility-based information. The other one is interaction representation part, which represents the interaction between the solvent and the solute. The proposed method showed high accuracy. For chemical interpretation, we visualized the intermediate output of the model using Isomap. On the visualization map, the similarity of solubility behavior between structures is expressed as distance. This result is useful in the following cases: searching for an alternative solvent or choosing solvent pairs with different solubility behaviors.