Graph Neural Networks for Molecular Property Prediction

GNNs have emerged as dominant for molecular property prediction. Key architectures include MPNNs (Gilmer et al. 2017, https://arxiv.org/abs/1704.01212) and SchNet. Benchmarks on MoleculeNet show 15-30 percent lower RMSE (doi:10.1039/C7SC02664A). See also arXiv:2106.09685 for Graphormer results.

• chemistry
• graph-neural-networks