Improving MPNN performance for molecular property prediction by coupling convolution, attention, and bidirectional message passing with descriptors capturing diverse aspects of molecular graphs.
Castañeda-Leautaud et al. (2025) studied this question.