Message Passing Networks for Molecules with Tetrahedral Chirality
This work tackles the problem of incorporating stereochemistry into molecular representations for drug discovery, which is a critical challenge for accurately predicting molecular properties.
This paper addresses the limitation of message passing neural networks (MPNNs) in handling molecular stereochemistry, specifically tetrahedral chirality. The authors developed two custom aggregation functions for MPNNs and evaluated them on synthetic data and a protein-ligand docking dataset, observing modest improvements over a baseline sum aggregator.
Molecules with identical graph connectivity can exhibit different physical and biological properties if they exhibit stereochemistry-a spatial structural characteristic. However, modern neural architectures designed for learning structure-property relationships from molecular structures treat molecules as graph-structured data and therefore are invariant to stereochemistry. Here, we develop two custom aggregation functions for message passing neural networks to learn properties of molecules with tetrahedral chirality, one common form of stereochemistry. We evaluate performance on synthetic data as well as a newly-proposed protein-ligand docking dataset with relevance to drug discovery. Results show modest improvements over a baseline sum aggregator, highlighting opportunities for further architecture development.