Path Complex Neural Network for Molecular Property Prediction

Abstract

Enormous power has been demonstrated by geometric deep learning (GDL) in molecular data analysis. However, there are still challenges in achieving high efficiency and expressivity in molecular representations, which are fundamental for the success of GDL. In this work, we introduce path complex neural network (PCNN) model for molecular property prediction. The essential idea is to use path complices to characterize various types of molecular interactions specified in molecular dynamic (MD) force field. We propose a path complex message-passing module to allow the communication of simplex features within/between different dimensions. Our model has been extensively validated on benchmark datasets and can achieve the state-of-the-art results.