Abstract

We present Symphony, an $E(3)$-equivariant autoregressive generative modelfor 3D molecular geometries that iteratively builds a molecule from molecularfragments. Existing autoregressive models such as G-SchNet and G-SphereNet formolecules utilize rotationally invariant features to respect the 3D symmetriesof molecules. In contrast, Symphony uses message-passing with higher-degree$E(3)$-equivariant features. This allows a novel representation of probabilitydistributions via spherical harmonic signals to efficiently model the 3Dgeometry of molecules. We show that Symphony is able to accurately generatesmall molecules from the QM9 dataset, outperforming existing autoregressivemodels and approaching the performance of diffusion models.

 

Quick Read (beta)