Bond Energy Assists Accurate Molecule Property Prediction

Molecule property prediction is a fundamental problem in many fields. To accurate and rapid prediction of molecules properties, molecule characterization and representation are key operations in the pretreatment stage. Generally, we represent molecule as graph based on different features such as element type, bond type, etc. The features we select could have the ability of enough representation and discrimination. However, in terms of bond embedding, one-hot coding is the most common processing method in the current research which means the features we used are discrete and could not distinguish different single or double bonds in a molecule. Here we add predicted bond energy feature as an extra chemical bond descriptor and compare three popular GNN models on two different datasets. The experiment shows that supplying additional unusual bond features—bond energy will improve the model performance significantly. The PDN model with bond energy has best performance among three model. Finally, we discuss the experimental result.